0% found this document useful (0 votes)

1K views

Data Compression

This is one of a series of subject guides publishe d by the University. Due to pressure of work the authors are unable to enter into any correspondence relating to, or arising from, the guide. If you have comments on this subject guide, favourabl e or unfavourable, please use the form at the back of this guide.

Uploaded by

Peter Rele

Available Formats

Download as DOC, PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

1K views

Data Compression

Uploaded by

Peter Rele

Available Formats

Download as DOC, PDF, TXT or read online on Scribd

You are on page 1/ 201

BSc in

Computing and Information Systems

Data compression
I. Pu 2004
2910 325

This guide was prepare d for the University of London by: I. Pu, PhD, BSc, Lecturer, Departmen t of Computing , Goldsmiths College, University of London. This is one of a series of subject guides publishe d by the University. We regre t that due to pressure of work the authors are unable to enter into any correspondence relating to, or arising from, the guide. If you have any comments on this subject guide, favourabl e or unfavourable , please use the form at the back of this guide.

The External Programme Publications Office University of London Stewart House 32 Russell Square London WC1B 5DN United Kingdom Web site: ww w.londonexternal.ac.uk Published by: University of London Press University of London 2004, reprinte d October 2005 (E7178) Printed by: Central Printing Service, University of London, England

Contents

Co ntents
0 Intr oduction Data compression . . . . . . . . . . . . . . . . . . . Aims and objectives of the subject . . . . . . . . . . Moti vation for studying the subject . . . . . . . . . Textbooks . . . . . . . . . . . . . . . . . . . . . . . Reading . . . . . . . . . . . . . . . . . . . . . Web addresses . . . . . . . . . . . . . . . . . . Prerequisites . . . . . . . . . . . . . . . . . . . . . . Study methods . . . . . . . . . . . . . . . . . . . . . Exercises, programming laboratory and courseworks Examination . . . . . . . . . . . . . . . . . . . . . . Subject guide . . . . . . . . . . . . . . . . . . . . . Activities . . . . . . . . . . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . . . . . . . . . . 1 Data compression Essential reading . . . . . . . . . . Further reading . . . . . . . . . . . Importance of data compression . . Brief history . . . . . . . . . . . . . Source data . . . . . . . . . . . . . Lossless and lossy data compression Lossless compression . . . . . Lossy compression . . . . . . . Main compression techniques . . . . Run-length coding . . . . . . . Qua ntisation . . . . . . . . . . Statistical coding . . . . . . . Dictionary-based coding . . . Transform-based coding . . . . Motion prediction . . . . . . . Compression problems . . . . . . . Algorithmic solutions . . . . . . . . Compression and decompression . . Compression performance . . . . . . Limits on lossless compression Learning outcomes . . . . . . . . . Activities . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . Sample examination questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii vii vii vii vii viii viii viii viii ix ix ix xi xi 1 1 1 1 1 2 2 2 2 3 3 3 3 3 4 4 4 4 5 5 6 6 8 8 8 9 9 9 9 i

2 Run-length algorithms Essential reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Run-length coding ideas . . . . . . . . . . . . . . . . . . . . . . . . . . Hardware data compression (HDC) . . . . . . . . . . . . . . . . . . . .

Data Compression

A simple HDC algorithm . . . Encoding . . . . . . . . . Decoding . . . . . . . . . Obser vation . . . . . . . . . . Learning outcomes . . . . . . Activities . . . . . . . . . . . . Laboratory . . . . . . . . . . . Sample examination questions

. . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9 10 10 11 11 12 12 12 13 13 13 13 13 14 16 17 17 17 19 20 21 21 21 23 23 23 23 24 25 25 26 26 27 27 27 28 29 30 31 32 32 32 35 35 35 35 35 35 36 36 38 38 38 39

3 Preliminaries Essential reading . . . . . . . . . . . Further reading . . . . . . . . . . . . Huffman coding . . . . . . . . . . . . Huffmans idea . . . . . . . . . . . . . Huffman encoding algorithm . . . . . Decoding algorithm . . . . . . . . . . Obser vation on Huffman coding . . . Shannon- Fano coding . . . . . . . . Shannon- Fano algorithm . . . . . . . Obser vation on Shannon- Fano coding Learning outcomes . . . . . . . . . . Activities . . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . . Sample examination questions . . . .

4 Coding symbolic data Essential reading . . . . . . . . . . . . Further reading . . . . . . . . . . . . . Compression algorithms . . . . . . . . . Symmetric and asymmetric compression Coding methods . . . . . . . . . . . . . Question of unique decodability . . . . Prefix and dangling suffix . . . . . . . . Prefix codes . . . . . . . . . . . . . . . Kraft-McMillan inequality . . . . . . . Some information theory . . . . . . . . Self-information . . . . . . . . . . Entropy . . . . . . . . . . . . . . . Optimum codes . . . . . . . . . . Scenario . . . . . . . . . . . . . . . . . Learning outcomes . . . . . . . . . . . Activities . . . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . . . Sample examination questions . . . . . 5 Huffman coding Essential reading . . . . . . . . . . . . Further reading . . . . . . . . . . . . . Static Huffman coding . . . . . . . . . Huffman algorithm . . . . . . . . . . . Building the binary tree . . . . . Canonical and minimum-variance Impleme ntation efficiency . . . . . . . . Obser vation . . . . . . . . . . . . . . . A problem in Huffman codes . . . . . . Extended Huffman coding . . . . . . . Learning outcomes . . . . . . . . . . . ii

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Contents

Activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sample examination questions . . . . . . . . . . . . . . . . . . . . . . .

40 40 40

6 Adapti ve Huffman coding 43 Essential reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Adaptive Huffman approach . . . . . . . . . . . . . . . . . . . . . . . . 43 Compressor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Encoding algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Function update tree . . . . . . . . . . . . . . . . . . . . . . . . 44 Decompressor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Decoding algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Function huffman next sym() . . . . . . . . . . . . . . . . . . . 44 . Function read unencoded sym() . . . . . . . . . . . . . . . . . . 44 Obser.vation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Learning outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Laboratory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Sample examination questions . . . . . . . . . . . . . . . . . . . . . . . 46 7 Arithmetic coding Essential reading . . . . . . . Further reading . . . . . . . . Arithmetic coding . . . . . . . Model . . . . . . . . . . . . . . Coder . . . . . . . . . . . . . . Encoding . . . . . . . . . . . . Unique-dec odability . . . Obser vation . . . . . . . Encoding main steps . . Defining an interval . . . Encoding algorithm . . . Obser vation . . . . . . . Decoding . . . . . . . . . . . . Renormalisation . . . . . . . . Coding a larger alpha bet . . . Effectiveness . . . . . . . . . . Learning outcomes . . . . . . Activities . . . . . . . . . . . . Laboratory . . . . . . . . . . . Sample examination questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 47 47 47 47 47 48 49 49 49 50 50 51 51 52 52 52 52 53 53 53 55 55 55 55 55 56 56 60 63 63 64 66 67 67 iii

8 Dictionary based compression Essential reading . . . . . . . . . . Further reading . . . . . . . . . . . Dictionary based compression . . . Popular algorithms . . . . . . . . . LZW coding . . . . . . . . . . . . . Encoding . . . . . . . . . . . . Decoding . . . . . . . . . . . . Obser vations . . . . . . . . . . LZ77 family . . . . . . . . . . . . . A typical compression step . . The decompression algorithm Obser vations . . . . . . . . . . LZ78 family . . . . . . . . . . . . .

Data Compression

One compression step . . Applications . . . . . . . . . . Highlight characteristics . . . Learning outcomes . . . . . . Activities . . . . . . . . . . . . Laboratory . . . . . . . . . . . Sample examination questions

. . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67 68 68 69 70 70 70 73 73 73 73 73 74 74 74 74 75 75 75 75 75 75 76 76 77 77 77 77 78 78 78 78 78 78 79 79 80 80 80 81 81 81 81 81 82 83 83 84 85 86 86 87 88 90 90

9 Image data Essential reading . . . . . . . . . . . . . . . . Further reading . . . . . . . . . . . . . . . . . Bitmap images . . . . . . . . . . . . . . . . . . Resolution . . . . . . . . . . . . . . . . . Displaying bitmap images . . . . . . . . . Vector graphics . . . . . . . . . . . . . . . . . Storing graphic components . . . . . . . Displaying vector graphic images . . . . . Difference between vector and bitmap graphics Combining vector graphics and bitmap images Rasterising . . . . . . . . . . . . . . . . . Vectorisation . . . . . . . . . . . . . . . . Colour . . . . . . . . . . . . . . . . . . . . . . RGB colour model . . . . . . . . . . . . . RGB represe ntation and colour depth . . LC represe ntation . . . . . . . . . . . . . Classifying images by colour . . . . . . . . . . Bi-level image . . . . . . . . . . . . . . . Gray-scale image . . . . . . . . . . . . . Colour image . . . . . . . . . . . . . . . . Classifying images by appearance . . . . . . . Continuous-tone image . . . . . . . . . . Discrete-tone image . . . . . . . . . . . . Cartoon-like image . . . . . . . . . . . . Colour depth and storage . . . . . . . . . . . . Image formats . . . . . . . . . . . . . . . . . . Obser vation . . . . . . . . . . . . . . . . . . . Learning outcomes . . . . . . . . . . . . . . . Activities . . . . . . . . . . . . . . . . . . . . . Laboratory . . . . . . . . . . . . . . . . . . . . Sample examination questions . . . . . . . . . 10 Image compression Essential reading . . . . . . . . . . . . . . Further reading . . . . . . . . . . . . . . . Lossless image compression . . . . . . . . . Bi-level image . . . . . . . . . . . . . Grayscale and colour images . . . . . . . . Reflected Gray codes (RGC) . . . . . Dividing a grayscale image . . . . . . Predicti ve encoding . . . . . . . . . . JPEG lossless coding . . . . . . . . . Lossy compression . . . . . . . . . . . . . . Distortion measure . . . . . . . . . . Progressive image compression . . . . Transforms . . . . . . . . . . . . . . . Karhunen-Loeve Transform (KLT) . . JPEG (Still) Image Compression Standard iv . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Contents

Learning outcomes Activities . . . . . . Laboratory . . . . . Sample examination

. . . . . . . . . . . . . . . . . . questions

. . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

90 91 91 92 93 93 93 93 93 93 94 94 94 94 96 96 97 97 98 98 99 99 99 99 100 100 100 100 101 101 101 101 102 102 102 103 103 104 104 105 105 105 105 105 106 106 106 107 109 115

11 Video compression Essential reading . . . . . . . . . . . . Further reading . . . . . . . . . . . . . Video systems . . . . . . . . . . . . . . Analog video . . . . . . . . . . . . Digital video . . . . . . . . . . . . Moving pictures . . . . . . . . . . . . . MPEG . . . . . . . . . . . . . . . . . . Basic principles . . . . . . . . . . . . . Temporal compression algorithms Group of pictures (GOP) . . . . . Motion estimation . . . . . . . . . . . . Work in different video formats . . . . Learning outcomes . . . . . . . . . . . Activities . . . . . . . . . . . . . . . . . Sample examination questions . . . . . 12 Audio compression Essential reading . . . . . . . . . . . . Further reading . . . . . . . . . . . . . Sound . . . . . . . . . . . . . . . . . . . Digital sound data . . . . . . . . . . . . Sampling . . . . . . . . . . . . . . Nyquist frequency . . . . . . . . . Digitisation . . . . . . . . . . . . . Audio data . . . . . . . . . . . . . . . . Speech compression . . . . . . . . . . . Pulse code modulation (ADPCM) Speech coders . . . . . . . . . . . Predicti ve approaches . . . . . . . Music compression . . . . . . . . . . . Streaming audio . . . . . . . . . . MIDI . . . . . . . . . . . . . . . . Learning outcomes . . . . . . . . . . . Activities . . . . . . . . . . . . . . . . . Sample examination questions . . . . . 13 Revision Revision materials . . . . Other references . . . . . Examination . . . . . . . Questions in examination Read questions . . . . . . Recommendation . . . . Importa nt topics . . . . . Good luck! . . . . . . . . A Sample examination B Exam solutions I paper II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

paper I

C Sample examination

131

Data Compression

D Exam solutions

137

E Sup port reading list 151 Text books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Web addresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

Chapter 0: Introduction

Chapter

Introduction
Data compression
Data compression is the science (and art) of represe nting information in a compact form. Having been the domain of a relatively small group of engineers and scientists, it is now ubiquitous. It has been one of the critical enabling technologies for the on-going digital multimedia revolution for decades. Without compression techniques, none of the ever-growing Internet, digital TV, mobile communication or increasing video communication would have been practical developments. Data compression is an active research area in computer science. By compressing data, we actually mean deriving techniques or, more specifically, designing efficient algorithms to: represe nt data in a less redunda nt fashion remove the redundancy in data impleme nt coding, including both encoding and decoding. The key approaches of data compression can be summarised as modelling + coding. Modelling is a process of constructing a knowledge system for performing compression. Coding includes the design of the code and product of the compact data form.

Aims and objecti ves of the sub ject

The subject aims to introduce you to the main issues in data compression and common compression techniques for text, audio, image and video data and to show you the significance of some compression technologies. The objectives of the subject are to: outline importa nt issues in data compression describe a variety of data compression techniques explain the techniques for compression of binary programmes, data, sound and image describe eleme ntary techniques for modelling data and the issues relating to modelling.

Moti vation for studying

the sub ject

You will broaden knowledge of compression techniques as well as the mathematical foundations of data compression, become aware of existing compression standards and some compression utilities available. You will also benefit from the development of transferable skills such as problem analysis and problem solving. You can also improve your programming skills by doing the laboratory work for this subject.

Text books
There are a limited number of books on Data compression available. No single book is completely satisfactory to be used as the textbook for the vii

Data Compression

subject. Therefore, instead of recommending one book for essential reading and a few books for further reading, we recommend chapters of some books for essential reading and chapters from some other books for further reading at the beginning of each chapter of this subject guide. An additional list of the books recommended for support and for historical background reading is atta ched in the reading list (Appendix E). Reading Salomon, David A Guide to Data Compression Methods. Springer, 2001) [ISBN 0-387-95260-8]. Wayner, Peter Compression Algorithms for Real Programmers . (London: Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapman, Nigel and Chapman, Jenny Digital Multimedia. John Wiley & Sons, 2000) [ISBN 0-471-98386-1]. (Chi chester: (London:

Sayood, Khalid Introduction to Data Compression. 2nd edition (San Diego: Morgan Kaufmann, 2000) [ISBN 1-55860-558-4]. Web addresses www.datacompression.com

Prerequisites
The prerequisites for this subject include a knowledge of elementary mathematics and basic algorithmics. You should review the main topics in mathematics, such as sets, basic probabili ty theory, basic computation on matrices and simple trigonometric functions (e.g. sin(x) and cos(x)), and topics in algorithms, such as data structures, storage and efficiency. Familiari ty with the elements of computer systems and networks is also desirable.

Study meth ods

As experts have predicted that more and more people in future will apply computing for multimedia, we recommend that you learn the importa nt principles and pay atte ntion to understanding the issues in the field of data compression. The experience could be very useful for your future career. We suggest and recommend highly the following specifically: 1. Spend two hours on revision or exercise for every hour of study on new material. 2. Use examples to increase your understanding of new concepts, issues and problems. 3. Always ask the question: Is there a better solution for the curre nt problem? 4. Use the Content pages to view the scope of the subject; use the Learning Outcomes at the end of each chapter and the Index pages for revision.

viii

Chapter 0: Introduction

Exercises,

programming

laboratory and course works

It is useful to have access to a computer so that you can actually impleme nt the algorithms lear nt for the subject. There is no restriction on the computer platform nor requireme nt of a specific procedural computer language. Examples of languages recommended include Java, C, C++ or even Pascal. Courseworks (issued separately every year) and tutorial or exercise/lab sheets (see Activities section and Sample examination questions at the end of each chapter) are set for you to check your understanding or to practice your programming skills using the theoretical knowledge gained from the course. The approach to impleme nting an algorithm can be different when done by different people, but it generally includes the following stages of work: 1. Analyse and understand the algorithm 2. Derive a general plan for impleme ntation 3. Develop the program 4. Test the correctness of the program 5. Comme nt on the limitations of the program. At the end, a full document should be written which includes a section for each of the above stages of work.

Examination
See the sample examination papers in Appendix A,C and solutions in Appendix B,D
1

The content in this subject guide will be examined in a two-hour-15-mi nute examinatio n1 . At the end of each chapter, there are sample examination questions for you to work on. You will normally be required to answer three out of five or four out of six questions. Each question often contains several subsections. These subsections may be classified as one of the following three types: Bookwork The answers to these questions can be found in the subject guide or in the main text book. Similar question The questions are similar to an example in the subject guide or the main text book. Unseen question You may have not seen these types of questions before but you should be able to answer them using the knowledge and experience gained from the subject. More information on how to prepare for your examination can be found in Chapter 13.

Su bject guide
The subject guide covers the main topics in the syllabus. It can be used as a reference which summarises, highlights and draws atte ntion to some importa nt points of the subject. The topics in the subject guide are equivalent to the material covered in a one term third- year module of BSc course in Mathematics, Computer Science, Internet Computing, or Computer Information Systems in London, which totals thir ty-three hours of lectures, ten hours of supervised laboratory work, and twenty hours of recommended ix

Data Compression

individual revisions or impleme ntation. The subject guide is for those stude nts who have completed all second year courses and have a successful experience of programming. This subject guide sets out a sequence to enable you to efficiently study the topics covered in the subject. It provides guidance for further reading, particularly in those areas which are not covered adequately in the course. It is unnecessary to read every textbook recommended in the subject guide. One or two books should be enough to enable individual topics to be studied in depth. One effective way to study the compression algorithms in the module is to trace the steps in each algorithm and attempt an example by yourself. Exercises and courseworks are good opportunities to help understanding. The sample examination paper at the end of the subject guide may also provide useful information about the type of questions you might expect in the examination. One thing the reader should always bear in mind is the fact that Data compression, like any other active research area in Computer Science, has kept evolving and has been updated, sometimes at an annoyingly rapid pace. Some of the descriptive information provided in any text will eventually become outdated. Hence you should try not to be surprised if you find different approaches, explanations or results among the books you read including this subject guide. The learning process requires the input of your own experiments and experience. Therefore, you are encouraged to, if possible, pursue articles in research journals, browse the relative web sites, read the latest versions of books, attend conferences or trade shows etc., and in general pay atte ntion to what is happening in the computing world. The contents of this subject guide are arranged as follows: Chapter 1 discusses essentials of Data compression, including a very brief histor y. Chapter 2 introduces an intuitive compression method: Run-length coding. Chapter 3 discusses the preliminaries of data compression, reviews the main idea of Huffman coding, and Shannon- Fano coding. Chapter 4 introduces the concepts of prefix codes. Chapter 5 discusses Huffman coding again, applying the information theory learnt, and derives an efficient impleme ntation of Huffman coding. Chapter 6 introduces adaptive Huffman coding. Chapter 7 studies issues of Arithmetic coding. Chapter 8 covers dictionary-based compression techniques. Chapter 9 discusses image data and explains related issues. Chapter 10 considers image compression techniques. Chapter 11 introduces video compression methods. Chapter 12 covers audio compression, and finally, Chapter 13 provides information on revision and examination. At the end of each chapter, there are Learning outcomes, Activities, laboratory questions and selected Sample examination questions. At the end of the subject guide, two sample examination papers and solutions from previous examinations in London can be found in the Appendix A-D.

Chapter 0: Introduction

Activities
1. Review your knowledge of one high level programming language of your choice, e.g. Java or C. 2. Review the following topics from your earlier studies of elementary mathematics and basic algorithmics: sets basic probabili ty theory basic computation on matrices basic trigonometric functions data structures, storage and efficiency. the elements of computer systems and networks

La boratory
1. Design and impleme nt a programme in Java (or in C, C++) which displays a set of English letters occurred in a given string (upper case only). For example, if the user types in a string AAABBEECEDE , your programme should display (A,B,E,C,D) . The user interface should be something like this: Please input a string: > AAABBEECEDE The letter set is: (A,B,E,C,D) 2. Write a method that takes a string (upper case only) as a parameter and that returns a histogram of the letters in the string. The ith element of the histogram should contain the number of ith character in the string alpha bet. For example, if the user types in a string AAABBEECEDEDEDDDE , then the string alpha bet is (A,B,E,C,D) . The output could be something like this: Please input a string: > AAABBEECEDEDED DDE The histogram is: A xxx B xx E xxxxxx Cx D xxxxx

Data Compression

Notes

xii

Chapter 1: Data compression

Chapter

Data compression
Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 1.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Introduction.

Importance of data compression

Data compression techniques is moti vated mainly by the need to improve efficiency of information processing. This includes improving the following main aspects in the digital domain: storage efficiency efficient usage of transmission bandwidth reduction of transmission time. Although the cost of storage and transmission bandwidth for digital data have dropped dramaticall y, the demand for increasing their capacity in many applications has been growing rapidly ever since. There are cases in which extra storage or extra bandwidth is difficult to achieve, if not impossible. Data compression as a means may make much more efficient use of existing resources with less cost. Active research on data compression can lead to innovative new products and help provide better services.

Brief history
Data compression can be viewed as the art of creating shorthand represe ntations for the data even today, but this process started as early as 1,000 BC. The short list below gives a brief survey of the historical milestones: 1000BC, Shorthand 1829, Braille code 1843, Morse code 1930 onwards, Analog compression 1950, Huffman codes 1975, Arithmetic coding 1977, Dictionary-based compression 1980s

early 80s, FAX mid-80s, Video conferencing, still images (JPEG), improved FAX standard (JBIG) late 80s, onward Motion video compression (MPEG) 1990s early 90s, Disk compression (sta cker) mid-90s, Satellite TV late 90s, Digital TV (HDTV), DVD, MP3 1

Data Compression

2000s Digital TV (HDTV), DVD, MP3

Source data
In this subject guide, the word data includes any digital information that can be processed in a computer, which includes text, voice, video, still images, audio and movies. The data before any compression (i.e. encoding) process is called the source data, or the source for short. Three common types of source data in the computer are text and (digital) image and sound. Text data is usually represe nted by ASCII code (or EBCDIC). Image data is represe nted often by a two-dimensional array of pixels in which each pixel is associated with its color code. Sound data is represe nted by a wave (periodic) function. In the application world, the source data to be compressed is likely to be so-called multimedia and can be a mixture of text, image and sound.

Lossless and lossy data compression

Data compression is simply a means for efficient digital represe ntation of a source of data such as text, image and the sound. The goal of data compression is to represe nt a source in digital form with as few bits as possible while meeting the minimum requireme nt of reconstruction. This goal is achieved by removing any redundancy presented in the source. There are two major families of compression techniques in terms of the possibility of reconstructing the original source. They are called Lossless and lossy compression. Lossless compression A compression approach is lossless only if it is possible to exactly reconstruct the original data from the compressed version. There is no loss of any information during the compressio n1 process. Lossless compression techniques are mostly applied to symbolic data such as character text, numeric data, computer source code and executable graphics and icons. Lossless compression techniques are also used when the original data of a source are so importa nt that we cannot afford to lose any details. For example, medical images, text and images preserved for legal reasons; some computer executable files, etc. Lossy compression A compression method is lossy compression only if it is not possible to reconstruct the original exactly from the compressed version. There are some insignifica nt details that may get lost during the process of compression. Appr oximate reconstruction may be very good in terms of the compression-ratio but usually it often requires a trade-off between the visual quality and the computation complexity (i.e. speed). Data such as multimedia images, video and audio are more easily compressed 2

This, when used as a general term, actually includes both comp ression and decompression process.

Chapter 1: Data compression

by lossy compression techniques.

Main compression

techniques

Data compression is often called coding due to the fact that its aim is to find a specific short (or shorter) way of represe nting data. Encoding and decoding are used to mean compression and decompression respectively. We outline some major compression algorithms below: Run-length coding Qua ntisation Statistical coding Dictionary-based coding Transform-based coding Motion prediction.

Run-length coding The idea of Run-length coding is to replace consecutively repeated symbols in a source with a code pair which consists of either the repeating symbol and the number of its occurrences, or sequence of non-re peating symbols. Example 1.1 String ABBBBBBBCC can be represented by Ar7 Br2 C, where r7 and r2 means 7 and 2 occurrences respectively. All the symbols are represe nted by an 8-bit ASCII codeword. Qua ntisation The basic idea of quantisation is to apply a certain computation to a set of data in order to achieve an appr oximation in a simpler form. Example 1.2 Consider storing a set of integers (7, 223, 15, 28, 64, 37, 145). Let x be an integer in the set. We have 7 x 223. Since 0 < x < 255 and 28 = 256, it needs 8 binary bits to represent each integer above. However, if we use a multiple, say 16, as a common divider to apply to each integer and round its value to the nearest integer, the above set becomes (0, 14, 1, 2, 4, 2, 9) after applying the computation x div 16. Now each integer can be stored in 4 bits, since the maximum number 14 is less than 24 = 16. Statistical coding The idea of statistical coding is to use statistical information to replace a fixed-size code of symbols by a, hopefully, shorter variable-sized code. Example 1.3 We can code the more frequently occurring symbols with fewer bits. The statistical information can be obtained by simply counting the frequency of each character in a file. Alternatively, we can simply use the probability of each character. Dictionary-based coding The dictionary approach consists of the following main steps: 3

Data Compression

1. read the file 2. find the frequently occurring sequences of symbols (FOSSs) 3. build up a dictionary of these FOSSs 4. associate each sequence with an index (usually a fixed length code) 5. replace the FOSS occurrences with the indices. Transform-based coding The transform-based approach models data by mathematical functions, usually by periodic functions such as cos(x) and applies mathematical rules to primarily diffuse data. The idea is to change a mathematical quantity such as a sequence of numbers 2 to another form with useful features. It is used mainly in lossy compression algorithms involving the following activities: analysing the signal (sound, picture etc.) decomposing it into frequency components making use of the limitations of human perception. M otion prediction Again, motion prediction techniques are lossy compression for sound and moving images. Here we replace objects (say, an 8 8 block of pixels) in frames with references to the same object (at a slightly different position) in the previous frame.

... or anything else

Compression

problems

In this course, we view data compression as algorithmic problems. We are mainly interested in compression algorithms for various types of data. There are two classes of compression problems of interest (Davisson and Gray 1976): Distortion-rate problem Given a constrai nt on transmitted data rate or storage capacity, the problem is to compress the source at, or below, this rate but at the highest fidelity possible. Compression in areas of voice mail, digital cellular mobile radio and video conferencing are examples of the distortion-rate problems. Rate-distortion problem Given the requireme nt to achieve a certain pre-specified fidelity, the problem is to meet the requireme nts with as few bits per second as possible. Compression in areas of CD-quali ty audio and motion-picture-quali ty video are examples of rate-distortion problems.

Algorithmic

solutions

In areas of data compression studies, we essentially need to analyse the characteristics of the data to be compressed and hope to deduce some patterns in order to achieve a compact represe ntation. This gives rise to a variety of data modelling and represe ntation techniques, which are at the heart of compression techniques. Therefore, there is no one size fits all solution for data compression problems. 4

Chapter 1: Data compression

Compression

and decompression

Due to the nature of data compression, any compression algorithm will not work unless a decompression approach is also provided. We may use the term compression algorithm to actually mean both compression algorithm and the decompression algorithm. In this subject, we sometimes do not discuss the decompression algorithm when the decompression process is obvious or can be easily derived from the compression process. However, you should always make sure that you know the decompression solutions. In many cases, the efficiency of the decompression algorithm is of more concern than that of the compression algorithm. For example, movies, photos, and audio data are often compressed once by the artist and then decompressed many times by millions of viewers. However, the efficiency of compression is sometimes more important. For example, programs may record audio or video files directly to computer storage.

Compression

performance

The performance of a compression algorithm can be measured by various criteria. It depends on what is our priority concern. In this subject guide, we are mainly concerned with the effect that a compression makes (i.e. the difference in size of the input file before the compression and the size of the output after the compression). It is difficult to measure the performance of a compression algorithm in general because its compression behaviour depends much on whether the data contains the right patterns that the algorithm looks for. The easiest way to measure the effect of a compression is to use the compression ratio. The aim is to measure the effect of a compression by the shrinkage of the size of the source in comparison with the size of the compressed version. There are several ways of measuring the compression effect: Compression ratio. This is simply the ratio of size.after.compression to size.before.compression or Compression ratio = Compression size.after.compression size.before.compression

factor. This is the reverse of compression ratio.

Compression factor =

size.before.compression size.after.compression

Saving perce ntage . This shows the shrinkage as a percentage. Saving percentage = size.before.compression size.after.compression % size.before.compression

Note: some books (e.g. Sayood(2000)) defines the compression ratio as our compression factor. Example 1.4 A source image file (pixels 256 256) with 65,536 bytes is compressed into a file with 16,384 bytes. The compression ratio is 1/4 and the compression factor is 4. The saving percentage is: 75% 5

Data Compression

In addition, the following criteria are normally of concern to the programmers: Overhead. Overhead is some amou nt of extra data added into the compressed version of the data. The overhead can be large sometimes although it is often much smaller than the space saved by compression. Efficiency This can be adopted from well established algorithm analysis techniques. For example, we use the big-O notation for the time efficiency and the storage requireme nt. However, compression algorithms behaviour can be very inconsiste nt but it is possible to use past empirical results. Compression time We normally consider the time for encoding and for decoding separatel y. In some applications, the decoding time is more importa nt than encoding. In other applications, they are equally importa nt. Entro py3 . If the compression algorithm is based on statistical results, then entropy can be used to help make a useful judgeme nt. Limits on lossless compression How far can we go with a lossless compression? What is the best compression we can achieve in a general case? The following two stateme nts may slightly surprise you: 1. No algorithm can compress all (possible) files, even by one byte. 2. No algorithm can compress even 1% of all (possible) files even by one byte. An informal reasoning for the above stateme nts can be found below: 1. Consider the compression of a big.file by a lossless compression algorithm called cmpres. If stateme nt 1 were not true, we could then effectively repeat the compression process to the source file. By effectively, we mean that the compression ratio is always < 1. This means that the size of the compressed file is reduced every time when running programme cmpres. So cmpres(cmpres(cmpres( cmpres(big.file ) ))), the output file after compression many times, would be of size 0. Now it would be impossible to losslessly reconstruct the original. 2. Compressing a file can be viewed as mapping the file to a different (hopefully shorter) file. Compressing a file of n bytes (in size) by at least 1 byte means mapping the file of n bytes to a file of n 1 bytes or fewer bytes. There are (28 )n = 256n files of n bytes and 256 n1 of n 1 bytes in total. This means that the proportion of the successful 1-to-1 mappings is only 256n1 /256n = 1/256 which is less than 1%.

We shall int roduce the concept of entropy later

Learning outcomes
On completion of your studies in this chapter, you should be able to: outline the brief history of Data compression 6

Chapter 1: Data compression

explain how to distinguish lossless data compression from lossy data compression outline the main compression approa ches measure the effect and efficiency of a data compression algorithm explain the limits of lossless compression.

Data Compression

Activities
1. Investigate what compression software is available on your computer system. 2. Suppose that you have compressed a file myfile using a compression utili ty available on your computer. What is the name of the compressed file? 3. Use a compression facility on your computer system to compress a text file called myfile containing the following text: This is a test. Suppose you get a compressed file called myfile.gz after compression. How would you measure the size of myfile and of myfile.gz? 4. Suppose the size of myfile.gz is 20 KB while the original file myfile is of size 40 KB. Compute the compression ratio, compression factor and saving percentage. 5. A compression process is often said to be negative if its compression ratio is greater than 1. Explain why negative compression is an inevitable consequence of a lossless compression.

La boratory
1. If you have access to a computer using Unix or Linux operating system, can you use compress or gzip command to compress a file? 2. If you have access to a PC with Windows, can you use WinZip to compress a file? 3. How would you recover your original file from a compressed file? 4. Can you use uncompress or gunzip command to recover the original file? 5. Impleme nt a method compressionRatio in Java which takes two integer argume nts sizeBeforeCompression and sizeAfterCompression and returns the compression ratio. See Activity 4 for example. 6. Similarly, impleme nt a method savingPercentage in Java which takes two integer argume nts sizeBeforeCompression and sizeAfterCompression and returns the saving percentage.

Sample examination

questions

1. Explain briefly the meanings of lossless compression and lossy compression. For each type of compression, give an example of an application, explaining why it is appropriate. 2. Explain why the following stateme nts are considered to be true in describing the absolute limits on lossless compression. No algorithm can compress all files, even by one byte. No algorithm can compress even 1% of all files, even by one byte.

Chapter 2: Run-length algorithms

Chapter

Run-length
Esse ntial reading

algorithms

Sayood, Khalid Introduction to Data Compression (Morgan Kaufmann, 1996) [ISBN 1-55860-346-8]. Chapter 6.8.1.

Run-length

coding ideas

A run-length algorithm assigns codewords to consecutive recurre nt symbols (called runs) instead of coding individual symbols. The main idea is to replace a number of consecutive repeating symbols by a short codeword unit containing three parts: a single symbol, a run-length count and an interpreting indicator. Example 2.1 String KKKKKKKKK, containing 9 consecutive repeating Ks, can be replaced by a short unit r9K consisting of the symbol r, 9 and K, where r represents repeating symbol, 9 means 9 times of occurrence and K indicates that this should be interpreted as symbol K (repeating 9 times). Run-length algorithms are very effective if the data source contains many runs of consecutive symbol. The symbols can be characters in a text file, 0s or 1s in a binary file or black-and-white pixels in an image. Although simple, run-length algorithms have been used well in practice. For example, the so-called HDC (Hardware Data Compression) algorithm, used by tape drives connected to IBM computer systems, and a similar algorithm used in the IBM SNA (System Network Architecture) standard for data communications are still in use today. We briefly introduce the HDC algorithm below.

Hard ware data compression

(HDC)

In this form of run-length coding, the coder replaces sequences of consecuti ve identical symbols with three elements: 1. a single symbol 2. a run-length count 3. an indicator that signifies how the symbol and count are to be interpreted.

A simple HDC algorithm

This uses only ASCII codes for: 1. the single symbols, and 2. a total of 123 control characters with a run-length count, including: repeating control characters: r2 , r3 , , r63 , and 9

Data Compression

non-re peating control characters: n1 , n2 , , n63 . Each ri , where i = 2 63, is followed by either another control character or a symbol. If the following symbol is another control character, ri (alone) signifies i repeating space characters (i.e. blanks). Otherwise, ri signifies that the symbol immediately after it repeats i times. Each ni , where i = 1 63 is followed by a sequence of i non- repeating symbols. Applying the following rules, it is easy to understand the outline of the encoding and decoding run-length algorithms below: Enc oding Repeat the following until the end of input file: Read the source (e.g. the input text) symbols sequentially and
1

i.e. a sequence of symbols.

1. if a string1 of i (i = 2 63) consecutive spaces is found, output a single control character ri 2. if a string of i (i = 3 63) consecutive symbols other than spaces is found, output two characters: ri followed by the repeating symbol 3. otherwise, identify a longest string of i = 1 63 non-re peating symbols, where there is no consecutive sequence of 2 spaces or of 3 other characters, and output the non-re peating control character ni followed by the string. Example 2.2 GGGBCDEFG55GHJKLM777777777777

can be compressed to r3 Gr6 n6 BCDEFG r2 n9 55tLMr12 7 Solution 1. The first 3 Gs are read and encoded by r3 G. 2. The next 6 spaces are found and encoded by r6 . 3. The non-re peating symbols BCDEFG are found and encoded by n6 BCDEFG . 4. The next 2 spaces are found and encoded by r2 . 5. The next 9 non-re peating symbols are found and encoded by n9 55GHJK LM. 6. The next 12 7s are found and encoded by r12 7. Therefore the encoded output is: r3 Gr6 n6 BCDEFG r2 n9 55tLMr12 7. Dec oding The decoding process is similar to the encoding and can be outlined as follows: Repeat the following until the end of input coded file: Read the codeword sequence sequentially and 10

Chapter 2: Run-length algorithms

1. if a ri is found, then check the next codeword. (a) if the codeword is a control character, output i spaces. (b) otherwise output i (ASCII codes of ) repeating symbols. 2. otherwise, output the next i non-re peating symbols.

Obser vation
1. It is not difficult to observe from a few examples that the performance of the HDC algorithm (as far as the compression ratio concerns) is:
2

It can be even better than entropy coding such as Huffman coding.

excellent2 when the data contains many runs of consecuti ve symbols poor when there are many segments of non-re peating symbols. Therefore, run-length algorithms are often used as a subroutine in other more sophisticated coding. 2. The decoding process of HDC is simpler than the encoding one 3

HDC is one of the so-called asymmetric coding methods. Please see page 24 for definition.

3. The HDC is non-adapti ve because the model remains unchanged during the coding process.

Learning outcomes
On completion of your studies in this chapter, you should be able to: state what a Run-length algorithm is explain how a Run-length algorithm works explain under what conditions a Run-length algorithm may work effectively explain, with an example, how the HDC algorithm works.

Data Compression

Activities
1. Apply the HDC (Hardware Data Compression) algorithm to the following sequence of symbols: kkkkghh5522777666abbbbcmmj## Show the compressed output and explain the meaning of each control symbol. 2. Explain how the compressed output from the above question can be reconstructed using the decompressing algorithm. 3. Provide an example of a source file on which the HDC algorithm would perform very badl y.

La boratory
1. Based on the outline of the simple HDC algorithm, derive your version of the HDC algorithm in pseudocode which allows an easy impleme ntation in Java (or C, C++). 2. Impleme nt your version of HDC algorithm in Java. Use MyHDC as the name of your main class/program. 3. Provide two source files good.source and bad.source, on which HDC would perform very well and very badly respectively. Indicate your definition of good and bad performance. [Hint] Define the input and the output of your (compression and decompression) algorithms first.

Sample examination

questions

1. Describe with an example how a Run-Length Coder works. 2. Apply the HDC (Hardware Data Compression) algorithm to the sequence: BCA1144330000EFGHHHH Demonstrate the compressed output and explain the meaning of each control symbol.

Chapter 3: Preliminaries

Chapter

Preliminaries
Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 2.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 1.

Huffman coding
Huffman coding is a successful compression method used originally for text compression. It assumes that each character is stored as a 8-bit ASCII code. You may have already come across Huffman coding in a programming course. If so, review the following two questions: 1. What property of the text does Huffman coding algorithm require in order to fulfil the compression? 2. What is the main idea of Huffman coding? Huffman coding works well on a text file for the following reasons:
Note: it is useful to distinguish the term codeword from the term cord although the two terms can be exchang eable sometimes. In this subject guide, a code consists of a number of codewords (see Example 3.1.)
1

Characters are represe nted normally by fixed-length codewords1 in computers. The codewords are often 8-bit long. Examples are ASCII code and EBCDIC code. Example 3.1 In ASCII code, codeword p1000001 represents character A; p1000010 B; p1000101 E , etc., where p is the parity bit. In any text, some characters occur far more frequently than others. For example, in English text, letters E,A,O,T are normally used much more frequently than J,Q,X. It is possible to construct a uniquely decodable code with variable codeword lengths. Our aim is to reduce the total number of bits in a sequence of 1s and 0s that represe nt the characters in a text. In other words, we want to reduce the average number of bits required for each symbol in the text.

Huffmans idea
Instead of using a fixed-length code for each symbol, Huffmans idea is to represe nt a frequently occurring character in a source with a shorter code and to represe nt a less frequently occurring one with a longer code. So for a text source of symbols with different frequencies, the total number of bits in this way of represe ntation is, hopefully, significantly reduced. That is to say, the number of bits required for each symbol on average is reduced. 13

Data Compression

Example E 5 A 5

3.2 Frequency of occurrence: O 5 T 3 J 3 Q 2 X 1

Suppose we find a code that follows Huffmans approach. For example, the most frequently occurring symbol E and A are assigned the shortest 2-bit codeword, and the lest frequently occurring symbol X is given a longer 4bit codeword, and so on, as below: E 10 A 11 O T J Q X 000 010 011 0010 0011

Then the total number of bits required to encode string EEETTJX is only 2 + 2 + 2 + 3 + 3 + 3 + 4 = 19 (bits). This is significantly fewer than 8 7 = 56 bits when using the normal 8-bit ASCII code.

Huffman enc oding algorithm

A frequency based coding scheme (algorithm) that follows Huffmans idea is called Huffman coding. Huffman coding is a simple algorithm that generates a set of variable-size codewords of the minimum average length. The algorithm for Huffman encoding involves the following steps: 1. Constructing a frequency table sorted in descending order. 2. Building a binary tree Carrying out iterations until completion of a complete binary tree: (a) Merge the last two items (which have the minimum frequencies) of the frequency table to form a new combined item with a sum frequency of the two. (b) Insert the combined item and update the frequency table. 3. Deriving Huffman tree Starting at the root, trace down to every leaf; mark 0 for a left branch and 1 for a right bran ch. 4. Generating Huffman code: Collecting the 0s and 1s for each path from the root to a leaf and assigning a 0-1 codeword for each symbol. We use the following example to show how the algorithm works: Example 3.3 Compress BILL BEATS BEN. (15 characters in total) using the Huffman approach. 1. Constructing the frequency table character frequency

B I L E A T S N SP(space) . 3 1 2 2 1 1 1 1 2 1 14

Chapter 3: Preliminaries

Sort the table in descending order: B L E SP I A T S N . 3 2 2 2 1 1 1 1 1 1 2. Building the binary tree

There are two stages in each step: (a) combine the last two items on the table (b) adjust the position of the combined item on the table so the table remains sorted. For our example, we do the following: (a) Combine B L E SP I A T S (N.) 3 2 2 2 1 1 1 1 2
Note: (N.) has the same frequency as L and E, but we have chosen to place it at the highest possible location - immediately after B (frequency 3).
2

Update 2 B (N.) 3 2

L E SP I A T S 2 2 2 1 1 1 1 L E SP I A 2 2 2 1 1 L E S 2 2 P 2 L 2

(b) B (TS) (N.) 3 2 2

(c) B (IA) (TS) (N.) 3 2 2 2

(d) (E SP) B (IA) (TS) (N.) 4 3 2 2 2 (e) ((N. ) 4 L)

(E SP) B ( A) (TS) 4 3 2 2

(f ) ((IA)(TS)) ((N.) L) (E SP) B 4 4 4 3 (g) ((E SP) B) ((IA)(TS)) ((N.) L) 7 4 4 (h) (((IA)(TS)) ((N.) L)) ((E SP) B) 8 7 (i) ((((IA)(TS)) ((N.) L)) ((E SP) B)) 15 The complete binary tree is: ((((IA)(TS)) ((N.) / (TS)) ((N.) / \ ((IA)(TS)) ((N.) / \ / (IA) (TS) (N.) /\ /\ /\ IA TS N. L)) ((E SP) B)) \ (((IA) L)) ((E SP) B) / \ L) (E SP) B \ /\ L E SP

Data Compression

3. Deriving

Huffman

tree L)) ((E SP) B)) \1 (((IA) L)) ((E SP) B) 0/ \1 L) (E SP) B \1 0/ \1 L E SP

((((IA)(TS)) ((N.) 0/ (TS)) ((N.) 0/ \1 ((IA)(TS)) ((N.) 0/ \1 0/ (IA) (TS) (N.) 0/ \1 0/ \1 0/ \1 IA TS N. 4. Generating Huffman

code

I A T S N . L E SP B 0000 0001 0010 0011 0100 0101 011 100 101 11 x 1 4 1 4 1 4 1 4 1 4 1 4 2 3 2 3 2 3 3 2

5. Saving perce ntage Comparison of the use of Huffman coding and the use of 8-bit ASCII or EBCDIC Coding: Huffman ASCII/EBCDIC Saving bits Percentage 48 120 72 60% 120 48 = 72 72/120 = 60%

Dec oding algorithm

The decoding process is based on the same Huffman tree. This involves the following types of operations: We read the coded message bit by bit. Starting from the root, we follow the bit value to traverse one edge down the the tree. If the curre nt bit is 0 we move to the left child, otherwise, to the right child. We repeat this process until we reach a leaf. If we reach a leaf, we will decode one character and re-start the traversal from the root. Repeat this read-move procedure until the end of the message. Example 3.4 Given a Huffman-coded message, 111000100101111000001001000111011100000110110101, what is the decoded message? ((((IA)(TS)) ((N.) L)) ((E SP) B)) 0/ \1 (((IA) (TS)) ((N.) L)) ((E SP) B) 0/ \1 0/ \1 ((IA)(TS)) ((N.) L) (E SP) B 0/ \1 0/ \1 0/ \1 (IA) (TS) N. L E SP 0/ \1 0/ \1 0/ \1 IA TS N . 16

Chapter 3: Preliminaries

After reading the first two 1s of the coded message, we reach the leaf B. Then the next 3 bits 100 lead us to the leaf E, and so on. Finally, we get the decoded message: BEN BEATS BILL.

Obser vation on Huffman coding

1. Huffman codes are not unique, for two reasons: (a) There are two ways to assign a 0 or 1 to an edge of the tree. In Example 3.3, we have chosen to assign 0 to the left edge and 1 for the right. However, it is possible to assign 0 to the right and 1 to the left. This would not make any difference to the compression ratio. (b) There are a number of different ways to insert a combined item into the frequency table. This leads to different binary trees. We have chosen in the same example to: i. make the item at the higher position the left child ii. insert the combined item on the frequency table at the highest possible position. 2. The Huffman tree built using our approach in the example tends to be more balanced than those built using other approaches. The code derived with our method in the example is called canoni cal minimum-varian ce Huffman code. The differences among the lengths of codewords in a canonical minimum-variance code turn out to be the minimum possible. 3. The frequency table can be replaced by a probabili ty table. In fact, it can be replaced by any appr oximate statistical data at the cost of losing some compression ratio. For example, we can apply a probabili ty table derived from a typical text file in English to any source data. 4. When the alpha bet is small, a fixed length (less than 8 bits) code can also be used to save bits.
3

i.e. the size is smaller or equal to 32.

Example 3.5 If the size of the alphabet set is not bigger than 323 , we can use five bits to code each character. This would give a saving percentage of 8 32 5 32 = 37.5%. 8 32

Shannon- Fano coding

This is another approach very similar to Huffman coding. In fact, it is the first well-known coding method. It was proposed by C. Shannon (Bell Labs) and R. M. Fano (MIT) in 1940. The Shannon- Fano coding algorithm also uses the probabili ty of each symbols occurrence to construct a code in which each codeword can be of different length. Codes for symbols with low probabilities are assigned more bits, and the codewords of various lengths can be uniquely decoded.

Shannon- Fano algorithm

Given a list of symbols, the algorithm involves the following steps: 1. Develop a frequency (or probabili ty) table 17

Data Compression

2. Sort the table according to frequency (the most freque nt one at the top) 3. Divide the table into 2 halves with similar frequency counts 4. Assign the upper half of the list a 0 and the lower half a 1 5. Recursively apply the step of division (2.) and assignme nt (3.) to the two halves, subdividing groups and adding bits to the codewords until each symbol has become a corresponding leaf on the tree. Example 3.6 Suppose the sorted frequency table below is drawn from a source. Derive the Shannon- Fano code. Symbol Frequency ----------------------A 15 B 7 C 6 D 6 E 5 Solution 1. First division: (a) Divide the table into two halves so the sum of the frequencies of each half are as close as possible. Symbol Frequency -------------------------A 15 22 B 7 11111111111111111111111111 First division C 6 17 D 6 E 5 (b) Assign one bit of the symbol (e.g. upper group 0s and the lower 1s). Symbol Frequency Code -------------------------------- A 15 0 B 7 0 11111111111111111111111111 First division C 6 1 D 6 1 E 5 1

Chapter 3: Preliminaries

2. Second division: Repeat the above recursively to each group. Symbol Frequency Code -------------------------------- A 15 00 22222222222222222222222222222 Second division B 7 01 11111111111111111111111111 First division C 6 10 22222222222222222222222222222 Second division D 6 11 E 5 11 3. Third division: Symbol Frequency Code -------------------------------- A 15 00 22222222222222222222222222222 Second division B 7 01 11111111111111111111111111 First division C 6 10 22222222222222222222222222222 Second division D 6 110 33333333333333333333333333333333 Third division E 5 111 4. So we have the following code (consisting of 5 codewords) when the recursive process ends: A B C D E 00 01 10 110 111

Obser vation on Shannon- Fano coding

1. It is not always easy to find the best division (see Step 3., the encoding algorithm). 2. The encoding process starts by assigning the most significant bits to each code and then works down the tree recursively until finished. It is as to construct a binary tree from the top. The following tree is for the example above: ABCD E 0/ \1 AB CDE 0/ \1 0/ A B C

\1 DE 0/ \1 D E

3. The decoding process is similar to Huffman decoding but the encoding process is not as simple and elegant as Huffman encoding.

Data Compression

Learning outcomes
On completion of your studies in this chapter, you should be able to: describe Huffman coding and Shannon- Fano coding explain why it is not always easy to impleme nt Shannon- Fano algorithm demonstrate the encoding and decoding process of Huffman and Shannon- Fano coding with examples.

Chapter 3: Preliminaries

Activities
1. Derive a Huffman code for string AAABEDBBTGGG. 2. Derive a Shannon- Fano code for the same string. 3. Provide an example to show step by step how Huffman decoding algorithm works. 4. Provide a similar example for the Shannon- Fano decoding algorithm.

La boratory
1. Derive a simple version of Huffman algorithm in pseudocode. 2. Impleme nt your version of Huffman algorithm in Java (or C, C++). 3. Similar to Lab2, provide two source files: the good and the bad. Explain what you mean by good or bad. 4. Impleme nt the Shannon- Fano algorithm. 5. Comme nt on the difference the Shannon- Fano and the Huffman algorithm.

Sample examination

questions

1. Derive step by step a canonical minimum-variance Huffman code for alpha bet {A, B, C, D, E, F}, given that the probabilities that each character occurs in all messages are as follows: Symbol ------B C D E F Probability ----------- A 0.3 0.2 0.2 0.1 0.1 0.1

2. Compute the average length of the Huffman code derived from the above question. 3. Given S = {A, B, C, D, E, F, G, H } and the symbols occurring probabilities 0.25, 0.2, 0.2, 0.18, 0.09, 0.05, 0.02, 0.01, construct a canonical minimum-variance Huffman code for this input.

Data Compression

Chapter 4: Coding symbolic data

Chapter

Coding symbolic data

Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 2.3-2.5.

Further

reading

Sayood, Khalid Introduction to Data Compression (Morgan Kaufmann, 1996) [ISBN 1-55860-346-8]. Chapter 2. In this chapter, we shall look more closely at the structure of compression algorithms in general. Starting with symbolic data compression, we apply the information theory to gain a better understanding of compression algorithms. Some conclusions we draw from this chapter may also be useful for multimedia data compression in later chapters.

Compression

algorithms

You will recall that in the Introduction, we said that data compression essentially consists of two types of work: modelling and coding. It is often useful to consciously consider the two entities of compression algorithms separatel y. The general model is the embodime nt of what the compression algorithm knows about the source domain. Every compression algorithm has to make use of some knowledge about its platform. Example 4.1 Consider the Huffman encoding algorithm. The model is based on the probability distribution of characters of a source text. The device that is used to fulfil the task of coding is usually called coder meaning encoder. Based on the model and some calculations, the coder is used to derive a code encode (compress) the input. Example 4.2 Consider the Huffman encoding algorithm again. The coder assigns shorter codes to the more frequent symbols and longer codes to infrequent ones. A similar structure applies to decoding algorithms. There is again a model and a decoder for any decoding algorithm. Conceptuall y, we can distinguish two types of compression algorithms, namely, static, or adaptive compression, based on whether the model structure may be updated during the process of compression or decompression. The model-coder structures can be seen clearly in diagrams Figure 4.1 and Figure 4.2: Static (non-adapti ve) system (Figure 4.1): This model remains unchanged during the compression or decompression process. 23

Data Compression

Model-C

Model-D

ABC etc.

Coder

010001100 etc.

Decoder

ABC etc.

compression

decompression

Figure 4.1: A static or non-adapti ve compression system

Adapti ve system (Figure 4.2): The model may be changed during the compression or decompression process according to the change of input (or feedback from the output). Some adaptive algorithms actually build the model based on the input starting from an empty model.

update Model-C

update Model-D

ABC etc.

Coder

010001100 etc.

Decoder

ABC etc.

compression

decompression

Figure 4.2: Adaptive compression system

In practice, the software or hardware for impleme nting the model is often a mixture of static and adaptive algorithms, for various reasons such as efficiency.

Symmetric

and asymmetric

compression

In some compression systems, the model for compression (Model-C in the figures) and that for decompression (Model-D) are identical. If they are identical, the compression system is called symmetri c, otherwise, it is said to be non-symmetri c. The compression using a symmetric system is called symmetric compression, and the compression using an asymmetric system is called asymmetric compression . 24

Chapter 4: Coding symbolic data

Coding meth ods

In terms of the length of codewords used before or after compression, compression algorithms can be classified into the following categories: 1. Fixed-to-fixe d: each symbol before compression is represe nted by a fixed number of bits (e.g. 8 bits in ASCII format) and is encoded as a sequence of bits of a fixed length after compression.
For ease of read, the symbol itself is used here instead of its ASCII codeword.
1

Example

4.3 A:00, B:01, C:10, D:11

2. Fixed-tovariable : each symbol before compression is represe nted by a fixed number of bits and is encoded as a sequence of bits of different length. Example 4.4 A:0; B:10; C:101; D:0101.

3. Variable-to-fixe d: a sequence of symbols represe nted in different number of bits before compression is encoded as a fixed-length sequence of bits. Example 4.5 ABCD:00; ABCDE:01; BC:11.

4. Variable-tovariable : a sequence of symbols represe nted in different number of bits before compression is encoded as a variable-length sequence of bits. Example Question 4.6 ABCD:0; ABCDE:01; BC:1; BBB:0001.

4.1 Which class does Huffman coding belong to?

Solution It belongs to the fixed-to-variable class. Why? Each symbol before compression is represe nted by a fixed length code, e.g. 8 bits in ASCII, and the codeword for each symbol after compression consists of different number of bits.

Question of unique dec odabili ty

The issue of unique decodability arises during the decompression process when a variable length code is used. Ideally, there is only one way to decode a sequence of bits consisting of codewords. However, when symbols are encoded by a variable-length code, there may be more than one way to identifying the codewords from the sequence of bits. Given a variable length code and a sequence of bits to decompress, the code is regarded as uniquely decodable if there is only one possible way to decode the bit sequence in terms of the codewords. Example 4.7 Given symbols A, B, C and D, we wish to encode them as follows: A:0; B:10; C:101; D:0101. Is this code uniquely decodable? The answer is No. Because an input such as 0101101010 can be decoded in more than one way, for example, as ACCAB or as DCAB.

Data Compression

However, for the example above, there is a solution if we introduce a new symbol to separate each codeword. For example, if we use a stop symbol /. We could then encode DDCAB as 0101/0101/101/0/10. At the decoding end, the sequence 0101/0101/101/0/10 will be easily decoded uniquel y. Unfortunatel y, the method above is too costly because of the extra symbol / introduced. Is there any alternati ve approach which allows us to uniquely decode a compressed message using a code with various length codewords? After all, how would we know whether a code with various length codewords is uniquely decodable? Well, one simple solution is to find another code which is a so-called prefix code such as (0,11,101,1001) for (A,B,C,D).

Prefix and dangling suffix

Let us look at some concepts first: Prefix: Consider two binary codewords w1 and w2 with lengths k and n bits respectively, where k < n. If the first k bits of w2 are identical to w1 , then w1 is called a prefix of w2 . Dangling Suffix: The remain of last n k bits of w2 is called the dangling suffix. Example 4.8 Suppose w1 = 010, w2 = 01011. Then the prefix of w2 is 010 and the suffix is 11.

Prefix codes
A prefix code is a code in which no codeword is a prefix to another codeword (Note: meaning prefix-free codes). This occurs when no codeword for one symbol is a prefix of the codeword for another symbol. Example 4.9 The code (1, 01, 001, 0000) is a prefix code since no codeword is a prefix of another codeword. The code (0, 10, 110, 1011) is not a prefix code since 10 is a prefix of 1011. Prefix codes are importa nt in terms of uniquely decodability for the following two main reasons (See Sayood(2000), section 2.4.3 for the proof ). 1. Prefix codes are uniquely decodable. This can be seen from the following informal reasoning: Example 4.10 Draw a 0-1 tree for each code above, and you will see the difference. For a prefix code, the codewords are only associated with the leaves. root 0/ \1 0/ \1 A 0/ \1 B 0/ C D root 0/ A \1 0/ \1 (B) 0/ \1 C \1 D

Chapter 4: Coding symbolic data

2. For any non-prefix code whose codeword lengths satisfy certain condition (see section Kraft-McMillan inequality below), we can always find a prefix code with the same codeword length distribution. Example 4.11 Consider code (0, 10, 110, 1011).

Kraft-McMillan

inequali ty

Theorem 4.1 Let C be a code with N codewords with lengths l1 , l2 , , lN . If C is uniquely decodable, then K (C ) =
N X

2li

1
i= 1

This inequality is known as the Kraft-McMillan inequali ty. (See Sayood(2000), section 2.4.3 for the proof ). P In N 2li 1, N is the number of codewords in a code, li is the length of i=1 the ith codeword. Example 4.12 Given an alphabet of 4 symbols (A, B, C, D), would it be possible to find a uniquely decodable code in which a codeword of length 2 is assigned to A, length 1 to B and C, and length 3 to D? Solution Here we have l1 = 2, l2 = l3 = 1, and l4 = 3.
4 X i= 1

1 1 1 1 + + + 3>1 22 2 2 2

Therefore, we cannot hope to find a uniquely decodable code in which the codewords are of these lengths. Example 4.13 If a code is a prefix code, what can we conclude about the lengths of the codewords? Solution Since prefix codes are uniquely decodable, they must satisfy the Kraft-McMillan Inequali ty.

Some information

theory

The information theory is based on mathematical concepts of probabili ty theory. The term information carries a sense of unpredictabili ty in transmitted messages. The information source can be represe nted by a set of event symbols (random variables) from which the information of each event can be measured by the surprise that the event may cause, and by the probabili ty rules that govern the emission of these symbols. The symbol set is frequently called the source alphabet, or alphabet for short. The number of elements in the set is called cardinality (|A| ). Self-information This is defined by the following mathematical formula: I (A) = logb P (A), 27

Data Compression

where A is an event, P (A) is the probabili ty that event A occurs. The logarithm base (i.e. b in the formula) may be in: unit bits: base 2 (used in the subject guide) unit nats: base e unit hartleys: base 10. The self-information of an event measures the amou nt of ones surprise evoked by the event. The negative logarithm logb P (A) can be written as logb 1 . P (A) Note that log(1) = 0, and that | log(P (A))| increases as P (A) decreases from 1 to 0. This supports our intuition from daily experience. For example, a low-probabili ty event tends to cause more surprise. Entro py The precise mathematical definition of this measure was given by Shannon. It is called entropy of the source which is associated with the experiments on the (random) event set. X X H = P (Ai )I (Ai ) = P (Ai ) logb P (Ai ), where source |A| = (A1 , . . . , AN ). The information content of a source is an importa nt attribute. Entropy describes the average amou nt of information converged per source symbol. This can also be thoug ht to measure the expected amou nt of surprise caused by the event. If the experime nt is to take out the symbols Ai from a source A, then the entropy is a measure of the minimum average number of binary symbols (bits) needed to encode the output of the source. Shannon showed that the best that a lossless symbolic compression scheme can do is to encode the output of a source with an average number of bits equal to the entropy of the source. Example 4.14 Consider the three questions below:

1. Given four symbols A, B, C and D, the symbols occur with an equal probability. What is the entropy of the distribution? 2. Suppose they occur with probabilities 0.5, 0.25, 0.125 and 0.125 respectively. What is the entropy associated with the event (experiment)? 3. Suppose the probabilities are 1,0,0,0. What is the entropy? Solution 1. The entropy is 1/4( log2 (1/4)) 4 = 2 bits 28

Chapter 4: Coding symbolic data

2. The entropy is 0.5 1 + 0.25 2 + 0.125 3 + 0.125 3 = 1.75 bits 3. The entropy is 0 bit.

Opti mum codes In information theory, the ratio of the entropy of a source to the average number of binary bits used to represe nt the source data is a measureme nt of the information efficiency of the source. In data compression, the ratio of the entropy of a source to the average length of the codewords of a code can be used to measure how successful the code is for compression. Here a source is usually described as an alpha bet = {s1 , , sn } and the next symbol chosen randomly from is si with probabili ty P r[si ] Pn = pi , i=1 pi = 1. Information Theory says that the best that a lossless symbolic compression scheme can do is to encode the output of a source with an average number of bits equal to the entropy of the source. We write done the entropy of the source: n X pi log2 1 = p log 1 + p log 1 + + p log1 1 2 n pi 2 2 p1 p2 p2 i= 1 n Using a variable length code to the symbols, li bits for si , the average number of bits is n X l = li pi = l1 p1 + l2 p2 + + ln pn
i=1

The code is optimum if the average length equals to the entropy. Let
n X i=1 n X

l i pi =

i=1 nX

1 pi log2 pi 1 pi log2 pi

and rewrite it as

n X i=1

pi li =

i=1

This means that it is optimal to encode each si with li = log2 pi bits, where 1 i n. It can be proved that: 1. the average length of any uniquely decodable code, e.g. prefix codes must be the entropy 2. the average length of a uniquely decodable code is equal to the entropy only when, for all i, li = log2 pi , where li is the length and pi is the probabili ty of codeword i 2 . 3. the average codeword length of the Huffman code for a source is greater and equal to the entropy of the source and less than the entropy plus 1. [Sayood(2000), section 3.2.3] 29

This can only happen if all probabilities are negative powers of 2 in Huffman codes, for li has to be an integer (in bits).

Data Compression

Scenario
What do the concepts such as variable length codes, average length of codewords and entropy mean in practice? Let us see the following scenario: Ann and Bob are extremely good friends but they live far away to each other. They are both very poor stude nts financially. Ann wants to send a shortest message to Bob in order to save her money. In fact, she decides to send to Bob a single symbol from their own secret alpha bet. Suppose that: the next symbol that Ann wants to send is randomly chosen with a known probabili ty Bob knows the alpha bet and the probabilities associated with the symbols. Ann knows that you have studied the Data compression so she asks you the importa nt questions in the following example: Example 4.15 Consider the three questions as below:

1. To minimise the average number of bits Ann uses to communi cate her symbol to Bob, should she assign a fixed length code or a variable length code to the symbols? 2. What is the average number of bits needed for Ann to communi cate her symbol to Bob? 3. What is meant by a 0 entropy? For example, what is meant if the probabilities associated with the alphabet are {0,0,1, , 0}? You give Ann the following: Solution 1. She should use variable length codes because she is likely to use some symbols more frequently than others. Using variable length codes can save bits hopefully. 2. Ann needs at least the P average number of bits that equal to the entropy n of the source. That is - i=1 pi log2 pi bits. 3. A 0 entropy means that the minimum average number of bits that Ann needs to send to Bob is zero. Probabili ty distribution {0,0,1, , 0} means that Ann will definitely send the third symbol in the alpha bet as the next symbol to Bob and Bob knows this. If Bob knows what Ann is going to say then she does not need to say anything, does she?!

Chapter 4: Coding symbolic data

Learning outcomes
On completion of your studies in this chapter, you should be able to: explain why modelling and coding are usually considered separately for compression algorithm design identify the model and the coder in a compression algorithm distinguish a static compression system by an adaptive one identify prefix codes demonstrate the relationship between prefix codes, KraftMcMillan inequality and the uniquely decodabili ty explain how entropy can be used to measure the code opti mum.

Data Compression

Activities
1. Given an alpha bet {a, b} with Pr[a]= 1/5 and Pr[b]= 4/5. Derive a canonical minimum variance Huffman code and compute: (a) the expected average length of the Huffman code (b) the entropy distribution of the Huffman code. 2. What is a prefix code? What can we conclude about the lengths of a prefix code? Provide an example to support your argume nt. 3. If a code is not a prefix code, can we conclude that it will not be uniquely decodable? Give reasons. 4. Determine whether the following codes are uniquely decodable: (a) {0,01,11,111} (b) {0,01,110,111} (c) {0,10,110,111} (d) {1,10,110,111}.

La boratory
1. Design and impleme nt a method entropy in Java which takes a set of probabili ty distribution as the argume nt and returns the entropy of the source. 2. Design and impleme nt a method averageLength in Java which takes two argume nts: (1) a set of length of a code; (2) the set of the probabili ty distribution of the codewords. It then returns the average length of the code.

Sample examination

questions

1. Describe briefly how each of the two classes of lossless compression algorithms, namely the adaptive and the non-adaptive, works in its model. Illustrate each with an appropriate example. 2. Determine whether the following codes for {A, B, C, D} are uniquely decodable. Give your reasons for each case. (a) {0, 10, 101, 0101} (b) {000, 001, 010, 011} (c) {00, 010, 011, 1} (d) {0, 001, 10, 010} 3. Lossless coding models may further be classified into three types, namely fixed-to-variable, variable-to-fixed, variable-to-variabl e. Describe briefly the way that each encoding algorithm works in its model. Identify one example of a well known algorithm for each model. 4. Derive step by step a canonical minimum-variance Huffman code for alpha bet {A, B, C, D, E, F}, given that the probabilities that each character occurs in all messages are as follows:

Chapter 4: Coding symbolic data

Symbol ------B C D E F

Probability ----------- A 0.3 0.2 0.2 0.1 0.1 0.1

Compare the average length of the Huffman code to the optimal length derived from the entropy distribution. Specify the unit of the codeword lengths used. Hi nt: log10 2 0.3; log10 0.3 0.52; log10 0.2 0.7; log10 0.1 = 1; 5. Determine whether the code {0, 10, 011, 110, 1111} is a prefix code and explain why. 6. If a code is a prefix code, what can we conclude about the lengths of the codewords? 7. Consider the alpha bet {A, B}. Suppose the probabili ty of A and B, Pr[A] and Pr[B] are 0.2 and 0.8 respectively. It has been claimed that even the best canonical minimum-variance Huffman coding is about 37% worse than its optimal binary code. Do you agree with this claim? If Yes, demonstrate how this result can be derived step by step. If No, show your result with good reasons. Hi nt: log10 2 0.3; log10 0.8 0.1; log10 0.2 0.7. 8. Following the above question, (a) derive the alpha bet that is expanded by grouping 2 symbols at a time; (b) derive the canonical Huffman code for this expanded alpha bet. (c) compute the expected average length of the Huffman code.

Data Compression

Chapter 5: Huffman coding

Chapter

Huffman
Esse ntial reading

coding

Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 2.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 1. In this chapter, we formally study the Huffman coding algorithms and apply the theory in Chapter 4.

Static Huffman coding

Assume that the fixed probability distribution is known to both the comp ressor and the decomp ressor.
1

The static Huffman coding1 is based on the following model: It includes a sequence of symbols from an alpha bet = {s1 , s2 , , sn }, where symbol si appears with a fixed probabili ty pi . The symbols are generated independently (which is called memoryless). This means that the probabili ty of finding si in the next character position is always pi , regardless of what went before.

Huffman algorithm
You may recall from the previous studies that the Huffman compression algorithm contains the following main operations: Build a binary tree where the leaves of the tree are the symbols to be coded. The edges of the tree are labelled by a 0 or 1. The decompression algorithm involves the operations where the codeword for a symbol is obtained by walking down from the root of the tree to the leaf for each symbol. Example 5.1 = {A, B, C, D, E, F } with probabilities 0.25, 0.2, 0.15, 0.15, 0.15, 0.1. Building the binary tree 1. Repeat the following until there is only one symbol in the alpha bet: (a) If there is one symbol, the tree is the root and the leaf. Otherwise, take two symbols in alpha bet s and s0 which have the lowest probabilities p and p0 . (b) Remove s and s0 from the alpha bet and add a new symbol [s, s0 ] with probabili ty p + p0 . Now the alpha bet has one fewer symbol than before. 2. Recursively construct the tree, starting from the root. 35

Data Compression

Canonical and mini mum- variance As we know: there can be items with equal probabilities the roles of 0 and 1 are interchangeable. To avoid this, we set the following rules: 1. A newly-created element is placed at the highest possible position in the list while keeping the list sorted 2. When combining two items, the one higher up in the list is assigned 0 and the lower down 1. The code that follows these rules is called canoni cal. By minimum-varian ce, we meant that the variation in the lengths of the codes is minimised. Huffman coding that follows these rules is called Canonical and Minimum-varian ce Huffman Coding.

Impleme ntation efficiency

The algorithm described earlier requires the system to: mai ntain the list of curre nt symbols in decreasing order of probabili ty, search the right place to insert the new symbols, which gives an O(n2 ) worst-case, where n is the number of symbols. One way to solve the efficiency problem is to do the following: Mai ntain 2 probabili ty (or frequency) lists: one (Li ) contains the original symbols in decreasing order of probabili ty. The other (Lc ) only contains super-symbols which are obtained by combining symbols, initially emp ty. A new combined symbol will always go to the front of list Lc (in O(1) time). Each time look at the ends of Li and Lc to determine which symbols next to combine. The next combination may be two supersymbols or two initial symbols or a super-symbol with an initial symbol. Example 5.2 Show how the algorithm works when the alphabet is {A,B,C,D,E,F,G,H,I, J} and the probabilities are (in %) 19, 17, 15, 13, 11, 9, 7, 5, 3, 1. 1. Li: A B C D E F 19 17 15 13 11 9 Lc: Empty 2. Li: A B C D E F 19 17 15 13 11 9 Lc: (IJ) 4 36 G 7 G 7 H 5 H 5 I 3 J 1

Chapter 5: Huffman coding

3. Li: A B C D E F 19 17 15 13 11 9 Lc: (H (IJ)) 9

G 7

4. Li: A B C D E 19 17 15 13 1 1 Lc: (FG) (H (IJ))) 16 9 5. Li: A B C D 19 17 15 13 Lc: (E (H (IJ))) (FG) 20 16 6. Li: A B 19 17 Lc: (CD) (E (H (IJ))) (FG) 28 20 16 7. Li: A 19 Lc: (B (FG)) (CD) (E (H (IJ))) 33 28 20 8. Li: Empty Lc: (A (E (H (IJ)))) (B (FG)) (CD) 39 33 28 Li: Empty 9. Lc: ((B (FG)) (CD)) (A (E (H (IJ)))) 61 39 Li: Empty 10. Lc: (((B (FG)) (CD)) (A (E (H (IJ))))) 10 0 Li: Empty Construct the binary tree recursively from the root: (((B (FG)) (CD)) (A (E (H (IJ))))) 0/ \1 ((B (FG)) (CD)) (A (E (H (IJ)))) 2. (((B (FG)) (CD)) (A (E (H (IJ))))) 0/ \1 0/ \1 0/ \1 (B (FG)) (CD) A (E (H (IJ))) 3. (((B (FG)) (CD)) (A (E (H (IJ))))) 0/ \1 0/ \1 0/ \1 0/ \1 0/\1 A 0/ \1 B (FG) C D E (H (IJ)) 4. (((B (FG)) (CD)) (A (E (H (IJ))))) 0/ \1 0/ \1 0/ \1 0/ \1 0/\1 A 0/ \1 B 0/\1 C D E 0/ \1 F G H 0/ \1 1. 37

Data Compression

So the code is: B 000 F 0010 G 0011 C 010 D 011 A 10 E 110 H 1110 H 1110 I 11110 J 11111

Obser vation
Note that: Huffman coding belongs to the fixed-to-variable coding meth od. Huffman codes are prefix codes (meaning prefix-free codes), so they are uniquely decodable. Huffman codes are optimal when probabilities of the symbols are all negative powers of 2, for li has to be an integer (in bits). Huffman codes are fragile for decoding: the entire file could be corrupted even if there is a 1-bit error.

A problem in Huffman codes

Here we mean the average number of bits of a code equals to the ent ropy.

Remember that Huffman codes meet the entropy bound only when all probabilities are powers of 2. What if the alpha bet is binary ( = {a, b})? The only optimal case2 is when pa = 1/2 and pb = 1/2. Hence, Huffman codes can be bad. Example 5.3 Consider a situation when pa = 0.8 and pb = 0.2.

Since Huffman coding needs to use 1 bit per symbol at least, to code the input, the Huffman codewords are 1 bit per symbol on average: 1 0.8 + 1 0.2 = 1 bits. However, the entropy of the distribution is (0.8 log2 0.8 + 0.2 log2 0.2) = 0.72 bits. This gives a gap of 1 0.72 = 0.28 bits. The performance of Huffman encoding algorithm is, therefore, 0.28/1 28% worse than optimal in this case.

Extended

Huffman coding

The idea of extended Huffman coding is to artificially increase the alpha bet size. For example, instead of assigning a codeword to every symbol, we derive a codeword for every 2 symbols. The following example shows how to achieve this: Example 5.4 Create a new alphabet = {aa, ab, ba, bb} extended from {a,b}. Let aa be A, ab for B, ba for C and bb for D. We now have an extend ed alphabet = {A, B, C, D}. The probability distribution can be calculated3 : P r[A] = P r[a] P r[a] = 0.64 P r[B] = P r[a] P r[b] = 0.16 38

Suppose that a or b occurs independently.

Chapter 5: Huffman coding

P r[C ] = P r[b] P r[a] = 0.16 P r[D] = P r[b] P r[b] = 0.04 We then follow the normal static Huffman encoding algorithm to derive the Huffman code for (See previous chapters for examples). The canonical minimum-variance code for this is A=0, B=11, C=100, D=101 . The average length is 1.56 bits for two symbols. The original output became 1.56/2 = 0.78 bits per symbol. This is only (0.78 0.72)/0.78 8% worse than optimal.

Learning outcomes
On completion of your studies of this chapter, you should be able to: explain how to improve the impleme ntation efficiency of static Huffman encoding algorithm by maintaining two sorted probabili ty lists illustrate some obvious weaknesses of Huffman coding describe how to narrow the gap between the minimum average number of binary bits of a code and the entropy using extended Huffman encoding meth od.

Data Compression

Activities
1. Given an alpha bet S = {A, B, C, D, E, F, G, H } of symbols with the probabilities 0.25, 0.2, 0.2, 0.18, 0.09, 0.05, 0.02, 0.01 respectively in the input, construct a canonical minimum-variance Huffman code for the symbols. 2. Construct a canonical minimum-variance code for the alpha bet A, B, C, D with probabilities 0.4, 0.3, 0.2 and 0.1. If the coded output is 101000001011, what was the input? 3. Given an alpha bet {a, b} with Pr[a] = 0.8 and Pr[b] = 0.2, determine the average code length if we group 3 symbols at a time. 4. Explain with an example how to improve the entropy of a code by grouping the alpha bet.

Sample examination

questions

1. Explain how the impleme ntation efficiency of a canonical minimum-variance Huffman coding algorithm can be improved by means of mai ntaining two frequency lists. 2. Derive step by step a canonical minimum-variance Huffman code for alpha bet {A, B, C, D, E, F} using the efficient impleme ntation approach, given that the probabilities that each character occurs in all messages are as follows: Symbol ------B C D E F Probability ----------- A 0.3 0.2 0.2 0.1 0.1 0.1

3. Consider the alpha bet {A, B}. Suppose the probabili ty of A and B, Pr[A] and Pr[B] are 0.2 and 0.8 respectively. It has been claimed that even the best canonical minimum-variance Huffman coding is about 37% worse than its optimal binary code. Do you agree with this claim? If Yes, demonstrate how this result can be derived step by step. If No, show your result with good reasons. Hi nt: log10 2 0.3; log10 0.8 0.1; log10 0.2 0.7. 4. For the above question, (a) derive the alpha bet that is expanded by grouping 2 symbols at a time; 40

Chapter 5: Huffman coding

(b) derive the canonical Huffman code for this expanded alpha bet; (c) compute the expected average length of the Huffman code.

Data Compression

Chapter 6: Adaptive Huffman coding

Chapter

Adapti ve Huffman coding

Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 5.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 1.5.

Adapti ve Huffman approa ch

The main characteristics of this approach include the following: The model calculates the probabilities of symbols as it goes along. For discussion convenience, we use the term weights instead of probabilities. The weight of a symbol is defined as the number of times that symbol has been seen (i.e. the frequency of its occurrence) so far. The output of the encoding (or decoding) process is generated immediately each time after reading in a symbol. The codes include Huffman codes as well as some fix-length codes. In the encoding process, for example, the model outputs a fix-length codeword (e.g. ASCII code) if the input symbol has been seen for the first time. Otherwise, it outputs a Huffman codeword. Let the alpha bet be = { , 1 , 2 , , n }, and g(i ) be any fixedlength codeword for i (e.g. ASCII code). In order to indicate whether the next output codeword is a fix-length or a variable-length codeword, one special symbol (6 ) is defined as a shift key solely for communication between the compressor and decompressor.

Compressor
The compression algorithm mai ntains a subset of symbols S of some alpha bet (i.e. S )) that the model has seen so far. A Huffman code (i.e. the tree) for all the symbols in S is also mai ntained. The weight of is always 0. The weight of any other symbol in S is its frequency so far. Initially, S = { } and the Huffman tree has the single node of symbol . During the process, the Huffman tree is used to assign codes to the symbols in S and updated after each output. Let h(i ) be the curre nt Huffman code for i and DAG for the special symbol in the algorithm below. Enc oding algorithm The following encoding algorithm shows how the compressor works:

Data Compression

1. Initialise the Huffman tree T containing the only node DAG. 2. while (more characters remain) do s:= next_symbol_in_text(); if (s has been seen before) then output h(s) else output h(DAG) followed by g(s) T:= update_tree(T); end next symbol in text() is a function that reads one symbol from the input sequence. update tree is another function which does the following: Function update tree 1. If s is not in S, then add s to S; weight_s:=1; else weight_s:=weight_s + 1; 2. Recompute the Huffman tree for the new set of weights and/or symbols.

Decompressor
The decompression algorithm mai ntains a subset of real symbols S 0 that the system has seen so far. The weight of is always 0, and the weight of any other symbol is the frequency of occurrence so far in the decoded output. Initially, S 0 := { } . Dec oding algorithm 1. Initialise the Huffman tree T with single node DAG 2. while (more bits remain) do s:=huffman_next_sym(); if (s==DAG) then s:= read_unencoded_sym(); else output s; T:=update_tree(T); end Function huffman next sym() The function reads bits from the input until it reaches a leaf node and returns the symbol with which that leaf is labelled. 1. start at root of Huffman tree; 2. while (not reach leaf) do read_next_bit(); follow edge; end 3. return the symbol of leaf reached. Function read unencoded sym() The function simply reads the next unencoded symbol from the input. For example, if the original encoding was an ASCII code, then it would read the next 7 bits (or 8 bits including the parity bit). As you saw earlier (in section Compressor), the function update tree does the following: 44

Chapter 6: Adaptive Huffman coding

1. If s is not in S, then add s to S else add 1 to the weight of s; 2. Recomputes the Huffman tree for the new set of weights and/or symbols.

Obser vation
Huffman coding, either static or adaptive, has two disad vantages that remain unsolved: Problem 1 It is not optimal unless all probabilities are negative powers of 2. This means that there is a gap between the average number of bits and the entropy in most cases. Recall the particularly bad situation for binary alphabets. Although by grouping symbols and extending the alpha bet, one may come closer to the optimal, the blocking method requires a larger alpha bet to be handled. Sometimes, extended Huffman coding is not that effective at all. Problem 2 Despite the availabili ty of some clever methods for counting the frequency of each symbol reasonably quickly, it can be very slow when rebuilding the entire tree for each symbol. This is normally the case when the probabili ty distributions change rapidly with each symbol.

Learning outcomes
On completion of your studies in this chapter, you should be able to: describe how adaptive Huffman coding algorithms work with examples identify impleme ntation issues of Huffman coding algorithms explain the two problems (weaknesses) of the Huffman coding algorithms in general.

Data Compression

Activities
1. Explain, with an example, how adaptive Huffman coding works. 2. Trace the adaptive Huffman coding algorithms to show how the following sequence of symbols is encoded and decoded: aaabbcddbbc cc

La boratory
1. Design and impleme nt a simple version of Adaptive Huffman encoding algorithm. 2. Design and impleme nt a simple version of Adaptive Huffman decoding algorithms.

Sample examination

questions

Chapter 7: Arithmetic coding

Chapter

Arithmetic
Esse ntial reading

coding

Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 4.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 1.7.

Arithmetic

coding

Arithmetic coding is the only replaceme nt whose benefits have been successfully demonstrated after 15 years of Huffman coding. It bypasses the idea of replacing an input symbol with a specific code. It replaces a stream of input symbols with a single floating- point output number. However, more bits are needed in the output number for longer, complex messages.
1

[0,1) means all real numbers 0 and < 1.

The output or Arithmetic coding is essentially a sequence of decimal digits, a fraction in the interval [0,1)1 . This single number can be uniquely decoded to create the exact stream of symbols that went into its construction. The arithmetic coding algorithm overcomes the problems in Huffman coding discussed at the end of Chapter 6: (for Problem 1) encodes a sequence of symbols at a time, instead of a single symbol. (for Problem 2) it outputs the coded version only after seeing the enti re input.

Model
The model is similar to the model for Huffman coding. Arithmetic coding based on an alpha bet and the probabili ty distribution of its symbols. Ideally, the probabilities are computed from the precise frequency counts of a source. This requires reading the entire source file before the encoding process. In practice, we can use an estimated or appr oximately fixed probabili ty distribution with the cost of slightly lower compression ratio. Similar to Huffman coding, the model can be a static or dynamic one depending on whether the probabili ty distribution is changed during a coding process.

Coder
We introduce a simple version of Arithmetic coder here to give a flavour of how it works. Suppose that both encoder and decoder know the length of the input sequence of symbols. Consider a simple case with a binary alpha bet {A, B}, and the probabili ty that the next input symbol is A is pA and pB . 47

Data Compression

Enc oding
Our essential goal is to assign an unique interval to each potential symbol sequence of a known length. After that, what we need to do is merely, from each interval, to select a suitable decimal number as the arithmetic code. The Arithmetic encoder reads a source of sequence of symbols from the above binary alpha bet one symbol at a time. Each time a new subinterval is derived depending on the probabili ty of the input symbol. This process, starting with interval [0,1), is iterated until the end of input symbol sequence. Then the arithmetic coder outputs a chosen decimal number within the final subi nterval for the entire input symbol sequence. Example 7.1 pA = 1/4, and pB = 3/4; and the symbol generation is memoryless 2 . Show the ideas of the simple version of Arithmetic coding for an input sequence containing 1. a single symbol 2. two symbols. Solution 1. Since pA is 1/4, we first divide the interval [0,1) into two subintervals of which the size is proportional to the probabili ty of each symbol, as shown in Figure 7.1. Using the convention that a real number < 1/4 represe nts A and a real number 1/4 represents B. Let A be .0 (output 0) and B as .5 (output 5).

By memoryless, we mean that the probability distribution for generating any symbol at any moment is the same as that of its original lifetime distribution.

0 A B

1/4

Figure 7.1: Encoding for one symbol. 2. Extending this to two symbols, we can see that pAA = pA pA = 1/16, pAB = pA pB = 3/16, pBA = pB pA = 3/16, and pBB = pB pB = 9/16. We divide each of the two subintervals further into two sub-subi ntervals of which the size is again proportional to the probabili ty of each symbol, as show in Figure 7.2. We then encode as follows: If the input is so far is A A ,i.e. input a A followed by another A ,then output a number <1/16=0.062 5; (code .0, i.e. output 0) If the input is A B then output a number >=1/16=0.0625 but <1/16+3/16=1/4=0.25; (code .1, output 1) If the input is B A then output a number >=1/4=0.25 but <1/4+3/16=7/16=0.4375; (code .3, output 3) 48

Chapter 7: Arithmetic coding

0 A AA 0 1/16 AB 1/4 BA 7/16 B BB

Figure 7.2: Encoding for two symbols. If the input is B B then output a number >=7/16=0.4375 but <7/16+9/16=1; (code .5, output 5).

Similarly, we can extend this to a sequence with 3 symbols: We can easily work out the probabili ty distribution {pAAA , pAAB , ...pBB B } and therefore decide which subinterval for each possible sequence. In theory, this subinterval division process can be repeated as many times as the number of input symbols requires. Unique-dec odabili ty We have shown in the previous section how it is possible to divide an interval starting from [0.1) according to the probabili ty of the input symbol each time and how to assign the final interval to every possible symbol sequence of a certain length. As we can see from the example that, in theory, there is no overlap among these subintervals for all possible symbol sequences. It is this overlap-free fact that makes arithmetic codes uniquely decodable. Obser vation We could choose, say, 0.05123456789 (<0.625) to encode AA but we do not, because 0.0 allows a one bit code, instead of a 11 bit one (for 0.05123456789). There are other equally good codes, For example, AB 0.2; BA 0.4; BB 0.6 (or 0.7, 0.8, 0.9) The code for A is the same as the code for AA. This is acceptable because the encoder and the decoder are both assumed to know the length of the input sequence. Enc oding main steps From previous examples, we know that the encoding process is essentially a process to derive the final interval following the main steps below: 1. Let the initial interval be [0,1) 2. Repeat the following until the end of input sequence: 49

Data Compression

(a) read the next symbol s in the input sequence (b) divide the curre nt interval into subintervals whose sizes are proportional to the symbols probabilities (c) make the subinterval for the sequence up to s the new curre nt interval 3. (When the entire input sequence has been processed in this way), output a decimal number within the curre nt interval. Defining an inter val There are usually two ways to describe an interval: 1. [L, H ], where variable L stores the lowest value of the interval range and H stores the highest. 2. [L, L + ], where variable L stores the lowest value of the interval range and stores the distance between the highest and the lowest value of the interval. We use the later approach for programming convenience. Enc oding algorithm We first look at a simple case for a small alpha bet containing only two symbols. Let the curre nt interval be [L, L + ) at each stage, and the output fraction x satisfy L x < L + . Initially, L=0 and = 1, this gives [0,1). If the next symbol could either be s1 or s2 with probabili ty p1 and p2 respectively, then we assign the intervals [L, L + p1 ) and [L + p1 , L + p1 + p2 ) and select the appropriate one. Note: L + p1 + p2 = L + (p1 + p2 ) = L +
In the algorithms below, < is used as an assignment operation (e.g. L < 0 means L is assigned a value 0)
3

Let d be in the algorithm below3 . 1. /* Initialise */ L <- 0 and d <- 1 2. Read next symbol 3. /* Update intervals */ If next symbol is s1, then leave L unchanged, and set d=d*p1 else L <- L + d*p1 and d <d*p2 4. If no more symbols left then output fraction from [L, L+d) else go to step 2. Example 7.2 For the example earlier, let s1 = A and s2 = B with probabilities p1 = 1/4 and p2 = 3/4. We encode ABA as follows: L = 0 and d = 1

Chapter 7: Arithmetic coding

read symbol A, leave L = 0 and set d = 1/4 read symbol B, set L = 1/16 and d = 3/16 read symbol A, leave L = 1/16 and set d = 3/64 Done, so choose a decimal number L = 1/16 = 0.0625 but < L + = 0.109375. For example, choose 0.1 and output 1. The following table shows arithmetic encoding algorithm that can be applied to encode all length 3 strings over the alpha bet {A,B} with pA = 1/4 and pB = 3/4. Seq AAA AAB ABA ABB BAA BAB BBA BBB Prob 1/64 3/64 3/64 9/64 3/64 9/64 9/64 27/64 Interval (Fraction) [0, 1/64) [1/64,4/64) [4/64,7/64) [7/64,16/64) [16/64,19/64) [19/64,28/64) [28/64,37/64) [37/64,1) Interval (Decimal) [0, .015625) [.015625, .0625) [.0625,.109375) [.109375, .25) [.25,.296875) [.296875,.4375) [.4375,.578125) [.578125, 1) Output 0 .02 .1 .2 .25 .3 .5 .6 log10 p 1.81 1.33 1.33 .85 1.33 .85 .85 .37

Obser vation In the algorithm and examples that we have discussed so far: A decimal system (base 10) is used for (description) convenience. The average length of the code is
1/64 + 6/64 + 3/64 + 9/64 + 6/64 + 9/64 + 9/64 + 27/64 = 70/64 = 1.09 digits

The information theoretic bound, i.e. the entropy, is 0.244 3 = 0.73 digit s. Huffman coding would require 3 digits for this case.

Dec oding
The decoding process is the inverse of the encoding process. Let p1 be the probabili ty for s1 and p2 be the probabili ty for s2. Given a string of digits which is a real number x in [0, 1), (d = ): 1. /* Initialise */ L <- 0 and d <- 1 2. Read x, 3. /* Computer the interval */ If x is a member of [L,L+d*p1) then output s1, leave L unchanged, and set d<-d*p1 else /* x is a member of [L+d*p1, L+d) */ then output s2, set L<- L+d*p1 and d<d*p2. 4. If the_number_of_decoded_symbols < the_required_number_of_symbols then go to step 2.

Data Compression

Renormalisation
Note the number of xs digits in the algorithms earlier can be quite large. On most computers, the value of would become zero rapidly (say after 300 symbols or so). This is a technique for dealing with an impleme ntation problem. The idea is to stop and L becoming zero rapidly by resetting the intervals. Example 7.3 Consider the Arithmetic encoding for baabaa .

After baa the interval is [.25,.296875), and we know that the first digit of the output must be 25. So output 25 and reset the interval to [.5,.96875); both become two to three times as big. This stops and L going to zero.

Coding a larger alpha bet

Arithmetic coding can be generalised to larger alpha bets. Example 7.4 Given 3 symbols A,B,C with probabilities 0.5, 0.3 and 0.2, the current allowed interval would be subdivided into 3 intervals according to the ratio 5:3:2. We would choose the new allowed interval among the three.

Effecti veness
The Arithmetic coding algorithms codes a sequence with probabili ty p with at most d log10 pe digits, and this is as close to the optimal as we will get in a digital world.

Learning outcomes
On completion of your studies in this chapter, you should be able to: explain the main ideas of Arithmetic coding describe, with an example of a small alpha bet {A,B}, how encoding and decoding Arithmetic algorithms work discuss the advantages of Arithmetic coding compared with Huffman coding explain the main problems in impleme ntation of Arithmetic coding.

Chapter 7: Arithmetic coding

Activities
1. What are the advantages of Arithmetic coding compared with the problems of Huffman coding? Use the following simple version of the encoding algorithm in the discussion. 1. /* Initialise */ L <- 0 and d <- 1 2. Read next symbol 3. /* Update intervals */ If next symbol is s1, then leave L unchanged, and set d=d*p1 else L <- L + d*p1 and d <d*p2 4. If no more symbols left then output fraction from [L, L+d) else go to step 2. 2. Demonstrate how to encode a sequence of 5 symbols, namely BABAB from the alpha bet {A,B} using the Arithmetic coding algorithm if pA = 1/5 and pB = 4/5. 3. Using the coding in the previous question as an example, explain how the Arithmetic decoding algorithm works. For example, explaining how to get the original BABAB back from the compressed result. 4. Show how to encode a sequence of 5 symbols from {A, B}, namely ABABB using arithmetic coding if pA = 1/5 and pB = 4/5. Would it be possible to have an coded output derived from a decimal value .24? Describe how the decoding algorithm works. 5. A sequence of 4 symbols from {A,B} was encoded using arithmetic coding. Assume pA = 1/5 and pB = 4/5. If the coded output is 0.24, derive the decoded output step by step.

La boratory
1. Impleme nt a simple version of the Arithmetic encoding algorithm for a binary alpha bet, e.g (A,B). 2. Impleme nt a simple version of the Arithmetic decoding algorithm for the same binary alpha bet (A,B).

Sample examination

questions

1. Describe briefly how Arithmetic Coding gets around the problems of Huffman Coding. You may use the following simple version of the encoding algorithm for discussion. 1. /* Initialise */ L <- 0 and d <- 1 2. Read next symbol 3. /* Update intervals */ If next symbol is s1, then leave L unchanged, and set d=d*p1 else L <- L + d*p1 and d <d*p2 4. If no more symbols left then output fraction from [L, L+d) else go to step 2. 2. Show how the Arithmetic coding is applied to code all length-2 strings over the alpha bet {A, B}. Suppose pA = 1/4 and pB = 3/4. You may like to summarise the code in the following format: 53

Data Compression

Sequence

Probability Output Fraction Decimal ------------------------------------------------??? ??? ??? ??? ... Interval

Interval

???

Chapter 8: Dictionary based compression

Chapter

Dictionary
Esse ntial reading

based compression

Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 3.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 2. In this chapter, we look at a set of algorithms based on a dictionary model.

Dictionary

based compression

These are techniques that incor porate the structure in the data in order to achieve a better compression. The main idea is to eliminate the redundancy of storing repetitive strings for words and phrases repeated within the text stream. The coder keeps a record of the most common words or phrases in a document and use their indices in the dictionary as output tokens. Ideally, the tokens are much shorter in comparison with the words or phrases themsel ves and the words and phrases are frequently repeated in the document. A list of commonly occurring patterns are built within the coder, including: front-end: this involves the selection of strings to be coded dictionary: this is used to match the strings to short codewords, such as indices, small integers or small pointers. Both compressors and decompressors have to mai ntain a dictionary. The dictionary-based algorithms are normally much faster than entropy-based ones, for they read the input as characters rather than as streams of bits. Dictionary based approaches can be
1

Note that there can be static dictionaries.

adapti ve1 variable-to-fixed in basic form fixed-to-variable in impleme ntation variable-tovariable overall. Applications include UNIX compress .

Popular algorithms
The most popular dictionary compression algorithms are: LZ77 (also known as LZ1) LZ78 (LZ2) LZW

Data Compression

The algorithms are named after the authors Abraham Lempel and Jakob Ziv who published the papers in 1977 and 1978. A popular varia nt of LZ78, known as (basic) LZW was published by Terry Welch in 1984. There are numerous variants of LZ77 and LZ78/LZW. We only, however, outline each of these algorithms here.

LZW coding
People often found that LZW algorithms are easier to understand and are the most popular ones. We hence study LZW coding first. Enc oding The dictionary usually contains 256 entries (e.g.ASCII codes) of single characters initiall y2 . The encoding algorithm is: 1. w o rd = ; 2. while not end_of_file x=read_next_character ; if word+x is in the dictionary /* where + means concatenate */ word=wor d+x else output the dictionary index for word; add word+x to the dictionary; word= x; 3. /* now end_of_file is true */ output the dictionary index for word; Example 8.1 Trace the operations of the LZW algorithm on the input string ACBBAAC. Suppose the first 256 places in the dictionary have been filled initially with symbols as below: Dictionary: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ABCDEFGHIJ K L M N O P Q R S T 21 22 23 25 27 ... U V W X Y Z space ... 24 26 Solution 1. Initial step: word: 2. Symbols to read: ACBBA AC read next_character x: A (in dictionary) word+x: A word : A 25 9 6

In the algorithms below, words between / * and * / signs are comments.

Chapter 8: Dictionary based compression

3. Symbols to read: CBBA AC read next_character x: C word+x: AC (not in dictionary) output: 1 new entry of the dictionary: 257 AC word : C 4. Symbols to read: BBAAC read next_character x: B word+x: CB (not in dictionary) output: 3 new entry of the dictionary: 258 CB word : B 5. Symbols to read: BAAC read next_character x: B word+x: BB (not in the dictionary) output: 2 new entry of the dictionary: 259 BB word : B 6. Symbols to read: AAC read next_character x: A word+x: BA (not in the dictionary) output: 2 new entry of the dictionary: 260 BA word : A 7. Symbols to read: AC read next_character x: A word+x: AA (not in the dictionary) output: 1 new entry of the dictionary: 261 AA word : A 8. Symbols to read: C read next_character x: C word+x: AC (in dictionary) word : AC 9. Symbols to read: none read next_character x: end_of_file output: 257 So the total output (the compressed file) is: 1 3 2 2 1 257. 57

Data Compression

The new entries of the dictionary are: 257 258 259 260 261 AC CB BB BA AA

Example Solution

8.2 Encode AAABAABBBB by tracing the LZW algorithm.

Suppose the first 256 places in the dictionary have been filled initially with symbols as below: Dictionary: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ABCDEFGHIJ K L M N O P Q R S T 21 22 23 24 25 26 27 ... U V W X Y Z space ... 1. Initial step: word: We show the actions taken and the variable values updated by the encoding algorithm on completion of each iteration of step 2. as follows: 2. Symbols to read: AAABAABBBB Input next_character (x): A word+x: A word : A 3. Symbols to read: AABAABBBB Input next_character (x): A word+x: AA Output token: 1 Dictionary (new entries): 257 A A word : A 4. Symbols to read: ABAABBBB Input next_character (x): A word+x: AA word : AA 5. Symbols to read: BAABBBB Input next_character (x): B word+x: AAB Output token: 257 Dictionary (new entries): 257 258 AA AAB 256 9

word : B 58

Chapter 8: Dictionary based compression

6. Symbols to read: AABBBB Input next_character (x): A word+x: BA Output token: 2 Dictionary (new entries): 257 258 259 AA AAB BA word : A 7. Symbols to read: ABBBB Input next_character (x): A word+x: AA word : AA 8. Symbols to read: BBBB Input next_character (x): B word+x: AAB word : AAB 9. Symbols to read: BBB Input next_character (x): B word+x: AABB Output token: 258 Dictionary (new entries): 257 258 259 260 AA AAB BA AABB word : B 10. Symbols to read: BB Input next_character (x): B word+x: BB Output token: 2 Dictionary (new entries): 257 258 259 260 261 AA AAA BA AABB BB word : B 11. Symbols to read: B Input next_character (x): B word+x: BB word : BB 12. Symbols to read: (none) Input next_character (x): end_of_file /* goes to step 3. */ Output token: 261 So the compressed output for AAABAABBBB is 1 257 2 258 2 261 59

Data Compression

The new entries of the dictionary are: 257 258 259 260 261 AA AAB BA AABB BB

Dec oding Similarly, the decoder builds its own dictionary as it reads the tokens one by one from the compressed file. Note: I have made some corrections to the decoding algorithm in book Wayner(2000) and use element to mean the dictionary entry at index (token) x. The algorithm now is: 1. read a token x from the compressed file; 2. look up dictionary for element at x; output element word=element ; 3. while not end_of_compressed_file do read x; look up dictionary for element at x; if there is no entry yet for index x then element=word+first_char_of_word; output element; add word+first_char_of_element to the dictionary; word=eleme nt; 4. end Example 8.3 Show the operations of the decompress algorithm on the input tokens 1 3 2 2 1 2573 . Solution 1. (Initial step) Tokens to be read: 1 3 2 2 1 257 read token x: 1 element: A output: A word : A 2. Tokens to be read: 3 2 2 1 257 read token x: 3 element: C output: C add new entry of dictionary: 257 AC word : C 3. Tokens to be read: 2 2 1 257

This is the compressed result from Example 8.1.

Chapter 8: Dictionary based compression

read token x: 2 element: B output: B add new entry of dictionary: 258 CB word : B 4. Tokens to be read: 2 1 257 read token x: 2 element: B output: B add new entry of dictionary: 259 BB word : B 5. Tokens to be read: 1 257 read token x: 1 element: A output A add new entry of dictionary: 260 BA word : A 6. Tokens to be read: 257 read token x: 257 element: AC output AC add new entry of dictionary: 261 AA word : AC 7. Token to be read: (none) read token x: end_of_compressed_file (end) So the total output is ACBBAAC . The new entries of the dictionary are: 257 258 259 260 261 AC CB BB BA AA

This is the compressed

result from Example 8.2.

Example 8.4 Decode 1 257 2 258 2 2614 of by tracing the decording algorithm. Suppose again the first 256 places in the dictionary have been filled initially with symbols as below: 61

Data Compression

Dictionary: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ABCDEFGHIJ K L M N O P Q R S T 21 22 23 25 27 ... U Y 24 V W X 26 Z space ... Solution 1. Tokens to be read: 1 257 2 258 2 261 Initial steps: Read next_token (x): 1 output: A word: A 2. Tokens to be read: 257 2 258 2 261 (Iterations begin): Read next_token (x): 257 Look up the dictionary and find there is no entry yet for 257 element: AA /* element=word+first_char_of_word */ output element: AA Dictionary (new entries so far): 257 AA word: AA 3. Tokens to read: 2 258 2 261 Read next_token (x): 2 Look up the dictionary and output element: B Dictionary (new entries so far): 257 258 AA AAB word: B 4. Tokens to read: 258 2 261 Read next_token (x): 258 Look up the dictionary and output element: AAB Dictionary (new entries so far): 257 258 259 AA AAB BA word: AAB 5. Tokens to read: 2 261 Read next_token (x): 2 Look up the dictionary and output element: B Dictionary (new entries so far): 257 258 259 260 AA AAB BA AABB word: B 6. Tokens to read: 261 25 9 6

Chapter 8: Dictionary based compression

Read next_token (x): 261 Look up the dictionary and find no entry for 261 element: BB /* element=word+first_char_of_word */ output element: BB Dictionary (new entries so far): 257 258 259 260 261 AA AAB BA AABB BB word: BB 7. Tokens to read: (none) Read next_token (x): end_of_compressed_file (end) So the decoded message is AAABAABBBB . Dictionary (new entries) is : 257 258 259 260 261 AA AAB BA AABB BB

Obser vations 1. The compression algorithm and the decompression algorithm build an identical dictionary independently. The advantage of this is that the compression algorithm does not have to pass the dictionary to the decompressor. 2. The decompressor lags one symbol behind. As we can see from Example 8.1, it only adds AA to cell 261 after actually outputting the entire decompressed string ACBBAAC .

LZ77 family
In this method, the dictionary to use is a portion of the previously seen input file. A sliding window that can be shifted from left to right is mai ntained to define the dictionary part and to scan the input sequence of symbols. The window consists of two parts: History buffer (also known as the Search buffer); this contains a portion of the recently seen symbol sequence Lookahead buffer; this contains the next portion of the sequence to be encoded. The size of each buffer is fixed in advance. Let H be the size of the Search buffer and L be that of the Lookahead buffer. Figure 8.1 shows an example of this setting: a History buffer of size H = 16 bytes (1 byte for each symbol) and Lookahead buffer of size L = 12, where the sequence of symbols in (and before) the History buffer has been seen and compressed but the sequence in (and after) the Lookahead buffer is to be compressed. You can image that the window moves from left to right (or the 63

Data Compression

(or, <-- text moving)

window moving -->

dictionary based compression compresses data using a dictionary History Lookahead

Figure 8.1: History buffer and Lookahead buffer text sequence moves from right to left) from time to time during the compression process. In practical impleme ntation, the History buffer is some thousands of bytes long and the lookahead buffer is only tens of bytes long (Salomon 2002). A typical compression step Suppose that the History buffer has H bytes (characters) that have been seen and encoded; The Lookahead buffer has at most L characters, which have been seen but not yet encoded. 1. Read characters from the input until the Lookahead buffer is full
By prefix of a symbol sequence, we mean a subsequence of any length beginning with the first symbol of the sequence. By offset of j bytes, we mean the number of the symbols from the first matched symbol to the end of the History buffer inclusive is j.
6 5

2. Scan the History buffer from right to left searching for a match to some prefix (of the sequence of symbols)5 in the Lookahead buffer. If more than one match is found in the History buffer, take the longest and the first one from the right, i.e. the one located furthest on the right in the History buffer. 3. If a match of length k 2 characters is found (see Figure 8.2) with an offset of j bytes6 then output the pointer (j, k). Slide the History buffer window k characters to the right (i.e. move the first k characters in the Lookahead buffer into the end of the History buffer, and remove the leftmost k characters from History buffer). 4. If no match is found, or if the match is only 1 character long (see Figure 8.1), then output (0, ASCII(c)), where c is the next character in the Lookahead buffer. Slide History window 1 character to the right.
window moving --> Lookahead

History

dictionary based compression compresses data using a dictionary 8 12

Figure 8.2: A match is found with an offset length 12 and match length 8. Example 8.5 Starting with a History buffer of size 16 bytes which contains the characters asedtcompression (where t represents a space) and an empty Lookahead buffer, show the output and the final state of History buffer if the following characters are next in the input, for Lookahead buffer of size 12: tcompresses tdatatusingtatdictionary . Solution We refer to Figure 8.2 and trace what happens step by step following the LZ77 encoding algorithm. 64

Chapter 8: Dictionary based compression

1. Read input string into Lookahead buffer and the History (left) and Lookahead (right) buffer now contain the following: asedtcompression tcompresses t datatusingtatdictionary A match tcompress is found and we record the offset 12 (bytes) and the match length 9 (bytes). Output: (12,9). Slide the window 9 characters to the right: The History and Lookahead buffer now contain: ression tcompress estdatatusin gtatdictionary
In fact, es is found in 2 places but we only consider the first (rightest) match here.
7

2. A match es is found7 so we record the offset 3 and the match length 2. Output: (3,2). Slide the window 2 characters to the right: The History and Lookahead buffer now contain: ssiontcompresses tdatatusingt atdictionary 3. Only a one-symbol match t is found so: Output: the pointer (0,ASC II(t)). Move the window 1 place to the right: The History and Lookahead buffer now contain: siontcompresses t datatusingta tdictionary 4. No match is found so: Output: the pointer (0,ASC II(d)). Move the window 1 place to the right: The History and Lookahead buffer now contain: iontcompresses td atatusingtat dictionary 5. No match is found so Output the pointer (0, ASCII(a)). Move History 1 place to the right. The History (left) and Lookahead (right) buffer now contain: ontcompresses tda tatusingtatd ictionary 6. No match is found so Output: the pointer (0,ASC II(t)), and move History 1 place to the right. The History (left) and Lookahead (right) buffer now contain: ntcompresses tdat atusingtatdi ctionary 7. Only one symbol match a is found so Output the pointer (0, ASCII(a)), and move History 1 place to the right. The History (left) and Lookahead (right) buffer now contain: tcompresses tdata tusingtatdic tionary . and so on ...

Data Compression

The decompression algorithm Decoder also mai ntains a buffer of the same size as the encoders window. However, it is much simpler than the encoder because there is no mat ching problems to deal with. The decoder reads a token and decide whether the token represents a mat ch. If it is a match, the decoder will use the offset and the match length in the token to reconstruct the match. Otherwise, it outputs the ASCII code in the token. We use an example to show how this works. Example 8.6 Suppose the buffer contains the following decoded symbols: Dictionary basedtcompression . The decoder reads the following tokens one at each iteration: (12,9) (3,2) (0,ASCII( t)) (0,ASCII(d))(0, ASCII(a)) (0,ASCII(t))(0, ASCII(a)). Solution 1. Read the next token (12,9), which is a match. The decoder uses the offset (the first number in the token) to find the first symbol of the match and uses the match length (the second number in the token) to decide the number of symbols to copy. So we have the following process: (a) Count 12 from the end of the buffer and find the symbol t (b) Copy the next 9 symbols one by one (see the underlined characters) and update the buffer: D ictionary based t compression t Di ctionary basedtcompression tc Dic tionary basedtcompression tco Dict ionary basedtcompression tcom Dicti onary basedtcompression tcomp Dictio nary basedtcompression tcompr Diction ary basedtcompr essiontcompre Dictiona ry basedtcompre ssiontcompres Dictionar y basedtcompres siontcompress 2. Read (3,2), the buffer becomes: based tcompression tcompre sses 3. Read (0,ASCII( )), which is not a match and output t. t sedtcompression tcompres sest 4. Read (0,ASCII(d)) which is not a match so output d (i.e. add d , into the buffer). edtcompression tcompress estd 5. Read (0,ASCII(a)) which is not a match so output a. , dtcompression tcompresse stda

Chapter 8: Dictionary based compression

6. Read (0,ASCII(t)) which is not a match so output t. , tcompression tcompresses tdat 7. Read (0,ASCII(a)) which is not a match so output a. , compression tcompresses t data . and so on.

Obser vations The History buffer is initialised to have no characters in it. Hence the first few characters are coded as raw ASCII codes (and the overhead) as the History buffer fills up. This may result in expansion rather than compression at the beginning. In common with the adaptive Huffman algorithm, LZ77 is an adapti ve algorithm and may start with an empty model. The decompression algorithm builds up the same History buffer as the compression algorithm, and decodes pointers with reference to this History buffer. The decompression algorithm knows if the next character is a real pointer or a raw character by looking at the first component of the pointer. In LZ77, there is an importa nt design decision concerns the values of H and L to be made: 1. Choosing a large History buffer means it is most likely that matches are found, but the offsets will be larger. A smaller buffer means smaller pointers but less chance of finding a mat ch. 2. Choosing a small Lookahead buffer means a quick search for a prefix, but the size of the match found will be limited. The algorithm is asymmetric since compression is slower than decompression. Compression involves searching for a match, which is computationally intensive, but decompression only involves reading out values from the History buffer.

LZ78 family
This is a commonly used data structu re for strings and similar to trees.
8

Typical LZ78 compression algorithms use a trie8 to keep track all of the strings seen so far. The dictionary D contains a set of strings, which are numbered from 0 onwards using integers. The number corresponding to a string is called the token of the string. The output of the encoding algorithm is a sequence of tokens. Initially the dictionary D is normally loaded with all 256 single- character strings, and the token of each single- character string is simply its ASCII code. All subseque nt strings are given token numbers 256 or more. One compression step

Let x be the currently matched string. Initially x is emp ty. The encoding algorithm is:

Data Compression

c:=next_char(); while (xc in D) do (* xc == string x followed by character c*) begin x:=xc; c:=next_char(); end output token corresponding to x; add xc to D, and give it the next available token x:=c;

Applications
UNIX utility compress is a widely used LZW varia nt. The number of bits used for represe nting tokens is increased as needed graduall y. For example, when the token number reaches 255, all the tokens are coded using 9 bits, until the token number 511 (29 1) is reached. After that, 10 bits are used to code tokens and so on. When the dictionary is full (i.e. the token number reaches the limit), the algorithm stops adapting and only uses existing strings in the dictionar y. At this time, the compression performance is monitored, and the dictionary is rebuilt from scratch if the performance deteriorates significantly. The dictionary is represe nted using a trie data structure. GIF (Graphics Interchange Format) is a lossless image compression format introduced by CompuServe in 1987. Each pixel of the images is an index into a table that specifies a colour map. The colour table is allowed to be specified along with each image (or with a group of images sharing the map). The table forms an uncompressed prefix to the image file, and may specify up to 256 colour table entries each of 24 bits. The image is really a sequence of 256 different symbols and is compressed using the LZW algorithm. V.42bis compression is the ITU-T standard for data compression for telephone-line modems. Each modem has a pair of dictionaries, one for incoming data and one for outgoing data. The maximum dictionary size is negotiated between the calling and called modem as the connection is made. The minimum size is 512 tokens with a maximum of 6 characters per token. Those tokens that are used infrequently may be deleted from the dictionar y. The modem may switch to transmitting uncompressed data if it detects that compression is not happening (e.g. if the file to transmit has already been compressed). The modem may also request that the called modem discards the dictionar y, when a new file is to be transmitted.

Highlig ht characteristics
We have studied all the importa nt compression algorithms for compressing text. Some of them are static and others are adaptive. Some use fixed length 68

Chapter 8: Dictionary based compression

codes, others use variable length codes. The compression algorithms and decompression algorithms use the same model, others use different ones. We summarised these characteristics below: Algorithm Adaptive Symmetric Type -----------------------------------------------------------Run-length n y variable Huffman n y fixed to Adaptive y y fixed to Huffman Arithmetic y y variable LZ77 y n variable LZW y y variable

to fixed variable variable to variable to fixed to fixed

Learning outcomes
On completion of your studies in this chapter, you should be able to: explain the main ideas of dictionary based compression describe compression and decompression algorithms such as LZ77, LZW and LZ78 list and comment on the main impleme ntation issues for dictionary based compression algorithms.

Data Compression

Activities
1. Using string abbaacabbabbb# (where # represents the end of the string) as an example, trace the values of word, x and the dictionary in running the basic Lempel-Ziv-Welch encoding and decoding algorithms, where word is the accu mulated string and x is the character read in on each round. 2. Suppose the input for encoding is a string aabbaaaccdee. Demonstrate how the simplified version algorithms mentioned above work step by step. Suppose the History buffer is 7 characters long and the LookAHead buffer is 4 characters long. Illustrate how LZ77 compression and decompression algorithms work by analysing the buffer content on each main step. For LZ78 and LZW, show what the dictionaries built are look like at the completion of the input sequence in the following format. Dictionary address (in decimal) ----------------------------------0 1 2 3 ... Dictionary entry ?, ?, ?, ?, ? ? ? ?

3. Describe a simplified version of LZ77, LZ78 and LZW algorithms. Analyse the type of the simplified algorithms in terms of, for example, whether they are static or adaptive and fixed-to-variabl e, variable-to-fixed or variable-to-variabl e.

La boratory
1. Impleme nt a simple version of LZW encoding algorithm. 2. Impleme nt a simple version of LZW decoding algorithm.

Sample examination

questions

1. Explain why a Dictionary-based coding method such as LZ77 is said to be adaptive and to be variable-to-fixed in its basic form. 2. One simplified version of LZ78/LZW algorithm can be described as below, where n is a pointer to another location in the dictionary and c is a symbol drawn from the source alpha bet, and < n,c > can form a node of a linked list. Suppose that the pointer variable n also serves as the transmitted code word, which consists of 8 bits. Suppose also the 0 address entry in the dictionary is a NULL symbol. The dictionary is initialised by constructing /* A Linked-list LZ78/LZW algorithm */ -------------------------------------1. n:=0; Fetch next source symbol c; 2. If the ordered pair <n,c> is already in the dictionary then n:= dictionary address of entry <n,c>; else 70

Chapter 8: Dictionary based compression

begin transmit n /* as a code word to decoder ??? */ create new dictionary entry <n,c> at dictionary address m; m:=m +1; n:= dictionary address of entry <0,c>; end; 3. Return to Step 1. Suppose that a binary information source emits the sequence of symbols 11000101100101110001111 (That is 110 001 011 001 011 100 011 11 without spaces). Construct the encoding dictionary step by step and show, in the format below, the dictionary on completion of the above input sequence. Dictionary address (in decimal) ----------------------------------0 1 2 3 ... Dictionary entry ?, ?, ?, ?, ? ? ? ?

Data Compression

Chapter 9: Image data

Chapter

Image data
Esse ntial reading
Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 3.1-3.2.

Further

reading

Chapman, Nigel and Chapman, Jenny Digital Multimedia. (John Wiley & Sons, 2000) [ISBN 0-471-98386-1]. Chapter 3-6. The compression techniques we have discussed so far are essentially developed for text compression. In this chapter, we look at a different type of data, namely, digital images. Image data is special in the area of data compression in comparison to video and audio data because it was the first challenge to modern data compression where much larger files have to be processed often as routine, and it was the first type of data to which lossy compression methods can be applied. Why do image files tend to be large compared with text? Why does image data allow lossy compression to be used? We look at the digital images in order to answer these questions and the answers are importa nt for us to understand image compression discussed in next chapter.
1

i.e. bitmap images.

In the multimedia world today, two types of images are most commonly used: one is the so-called bitmap images1 and the other is vector graphics.

Bitmap images
These are also called photographic images for 2-dimensional pictures. A bitmap image is an array of pixel values. The data are the values of the pixels. Many bitmap images are created from digital devices such as scanners or digital cameras, or from programs such Painter which allows visual artists to paint images. Resolution A bitmap image is often measured by resolution which is a measure of how finely a device appr oximates continuous images using finite number of pixels. This is closely related to sampling rates. There are two common ways of describing resolution: dpi stands for dots per inch and represents the number of dots per unit length for data from devices such as scanners. pixel dimension is the number of pixels per frame for the video data from digital cameras. For TV systems, the PAL frame is 768 by 576 pixels and the NTSC frame is 640 by 480. For computer monitors, the common resolutions are 640 by 480 or 1024 by 768. 73

Data Compression

Displ aying bitmap images Since a bitmap image is in fact represe nted in a computer by an array of pixel values, there has to be a way of mapping each pixel value to the physical dots on a display screen. In the simplest case, each pixel corresponds one-to-one to the dots on the screen. In the general case, the physical size of the display of an image will depend on the device resolution. The pixel values are stored at different resolutions from the displayed image. Computations for scaling and clipping usually need to be performed to display a bitmap image. In general, we have the following formula to decide the image sign and scaling: imageDimension = pixelDimension/deviceResolution. where the deviceResolution is measured in dpi.

Vector graphics
The image in vector graphics is stored as a mathematical description of a collection of graphic components such as lines, curves, and shapes and so on. Vector graphics are normally more compact, scalable, resolution independent and easy to edit. They are very good for 3-D models on computer, usually built up using shapes or geometric objects that can easily be descri bed mathematicall y. Storing graphic com pone nts The graphic components are stored by their mathematical functions instead of pixels. For example, consider storing a line in the computer. Example 9.1 The mathemati cal function for a straight line is y = kx + b, where k is the slope and b the intercept. The function, instead of all the pixels, can be stored for the line. Since a geometric line is actually a finite segment of a line, we also need to store the two end points of the segment. In fact, it is sufficient to only store the coordinates of the two end points because we know the line function can be represe nted by two point coordinates. That is, given (x1 , y1 ) and (x2 , y2 ), we can easily write the function for the line: y = (y2 y1 )/(x2 x1 )x + (x2 y1 x1 y2 )/(x2 x1 ). Displ aying vector graphic images This would require some computation to be performed in order to interpret the data and generate an array of pixels to be displayed. The process of generating a pattern of pixels from a model is called rendering . If the image is a line, only two end points are stored. When the model is rendered for display, a pattern of pixels has to be generated to display the line with the two end points.

Chapter 9: Image data

Difference between vector and bitmap graphics

The differences between the two can be very obvious in terms of visual characteristics, but we are concerned more about the following issues: 1. The requireme nts of the computer system: a bitmap image must record the value of every pixel, but vector description takes much less space, so is more economical. 2. The size of bitmap images depends on the size of the image and the resolution but is independent of the complex of the image; while the size of vector graphics depends on the number of objects of which the image consists, it is independent of any resolution. 3. The approa ch of so-called painting programs produces bitmap images while that of drawing programs produces vector graphics. Very few packages offer support to both types of images. 4. The behaviour of the bitmap images and that of vector graphics are different when re-sized or scaled.

Co mbining vector graphics and bitmap images

The combination requires a transformation between vector graphics and bitmap images. Rasterising A raster is a predetermined pattern of scanning lines to provide substa ntially uniform coverage of a display area. Rasterising is the process of interpreting the vector description of graphics for a display area. . A vector graphic loses all the vector properties during the process, e.g. the individual shapes can no longer be selected or moved because they become pixels. Vectorisation This is a more complicated process of transforming pixels to vectors. The difficulties arise from the need to identify the boundary of a pixel image using the available forts of curves and lines and to colour them. The vector file tends to be much bigger than the pixels file before the vectorization.

Colour
As we know, colour is a subjective sensation experience of the human visual system and the brain. It is a complicated system. Nevertheless, according to simplified so-called tristimulus theory, in practice we can define a particular colour by giving the proportions of Red, Green and Blue light that the colour contains. RGB colour model This is a model for computing purposes. It comes from the idea of constructing a colour out of so-called additive primary colours (i.e. Red, Green and Blue or the RGB for short). Although there is no universally accepted standard for RGB, television and video industry does have a standard version of RGB colour derived in the Recommendation ITU-R BT.709 for High Definition TV (HDTV). Monitors are increasingly being built to follow the recommendation. 75

Data Compression

In the RGB model, we assume that all colours can in principle be represe nted as combinations of certain amounts of red, green and blue. Here the amou nt means the proportion of some standard of primary red, green and blue. Example 9.2 Suppose that the proportion is represented as percentages. (100%,0%,0%) then represents a colour of pure red, and (50%,0%,0%) a darker red (0%, 0%, 100%) a pure blue (0%, 0%, 0%) black . (100%, 100%, 100%) .and so on. In fact, there are two commonly used represe ntations of digital colours: they are so-called RGB representation and LC representation . RGB represe ntation and colour depth Since it is only the relative values of R, G and B that matters, we can actually choose any convenient value range, as long as the range provided distinguishes enough different values. In a common RGB represe ntation, we use 1 byte (8 bits) to describe the lightness of each of the three primary colours, which gives a range of [0,255] (or [1,256]). Three bytes (8 3 = 24 bits) describe each pixel of a colour image. In this way, there are 2563 (i.e. 16,777,216) different colours can be represe nted in this way. The number of bits used to hold a colour value is often called colour depth. For example, if we use 24 bits to represe nt one pixel, then the colour depth is 24. We sometimes also say 24 bit colour to refer to a colour with 24 bit colour depth. Example colours: 9.3 Some colours from the Linux colour database are 24 bit

R G B colour G B colour -----------------------------------------------------25 25 255 white 25 25 250 snow 5 5 128 navy blue 5 0 255 ghost 0 0 24 24 8 8 213 white 0 0 255 blue1 25 23 papaya 5 9 225 whip rose 0 25 0 green 25 22 misty 5 8 ... 5 LC represe ntation This is another common represe ntation which is based on luminance (Y) and chrominance (C) values. Luminance Y reflects the brightness, and by itself gives a gray-scale version of the image. The approach is based on colour differences. The idea is to separate the brightness information of an image from its colour. By separating brightness and colour, it is possible to transmit a picture in a way that the colour information is undetected by a black and white receiver, which can simply treat the brightness as a monochrome signal. 76

Chapter 9: Image data

A formula is used which has been empirically determined for the best gray-scale likeness of a colour image. Y = 0.299R + 0.587G + 0.114B The chrominance (colour) components provide the additional information needed to convert the gray-scale image to a colour image. These components are represe nted by two values Cb and Cr given by Cb = B Y and Cr = R Y . This technique was initially used during the development from blackand-white to colour TV. However, it turns out to have some advantages. For example, since the human eye is much better at distinguishing subtle differences in brightness than subtle differences in colour, we can compress the Y component with greater accuracy (lower compression ratio), and make up for it by compressing the Cb and Cr components with less accuracy (higher compression ratio).

Classifying images by colour

For compression purposes, we can classify the images for compression by number of colours used. Three approaches are available, namely bi-level image, gray-scale image, colour image. Bi-le vel image This is actually the image with 1 bit colour. One bit allows us to distinguish two different colours. Images are captured by scanners using only two intensity levels, one is information and the other background. Applications include: Text, line drawings or illustrations FAX documents for transmission Sometimes photographs with shades of gray. Gr ay-scale image This is actually the image with a 8-bit colour depth. The images are captured by scanners using multiple intensity levels to record shadings between black and white. We use 8 bits to represe nt one pixel, and hold one colour value to provide 256 different shades of grey. Gray scale images are appropriate for medical images as well as black-and-white photos. Colour image Colour images are captured by scanners, using multiple intensity levels and filtering to capture the brightness levels for each of primary colours, R, G and B (Red, Green, Blue). Some computer systems use 16 bits to hold colour values. In this case, either 1 bit is left unused or different numbers of bits are assigned to R, G and B. If it is the latter, then we usually assign 5 bits for R and B, but 6 bits for G. This allocation is due to the fact that the human eye is more sensitive to green light than to red and blue. 77

Data Compression

Most common colour images are of a 24 bits colour depth. Although 24 bits are sufficient to represe nt more colours than the eye can distinguish, higher colour depths such as 30, 36 or even 48 bits are increasingly used, especially by scanners.

Classifying images by ap pearance

Images can be classified by their appearances which are caused by the way in which colours distribute, namely continuous-tone image, discrete-ton image and cartoon-like image. Co ntinuous-tone image This type of image is a relatively natural image which may contain areas with colours. The colours seem to vary continuously as the eye moves along the picture area. This is because the image usually has many similar colours or grayscales and the eye cannot easily distinguish the tiny changes of colour when the adjace nt pixels differ, say, by one unit. Examples of this type of image are the photographs taken by digital cameras or the photographs or paintings scanned in by scanners. Discrete-tone image Alternati ve names are graphical image or synthetic image. This type of image is a relatively artificial image. There is usually no noise and blurring as there is in a natural image. Adjace nt pixels in a discrete-tone image are often either identical or vary significantly in value. It is possible for a discrete-tone image to contain many repeated characters or patterns. Examples of this type of images are photographs of artificial objects, a page of text, a chart and the contents of a computer screen. Cart oon-li ke image This type of image may consist of uniform colour areas but adjace nt areas may have very different colours. The uniform colour areas in cart oon-like images are regarded as good features for compression.

Colour depth and storage

Colour depth determines the size of the bitmap image: each pixel requires 24 bits for 24 bit colour, but just a single bit for 1 bit colour. Hence, if the colour depth is reduced from 24 to 8, the size of a bitmap image will decrease by a factor of 3 (ignoring any fixed-size housekeeping information).

Image formats
There are lot of image software available these days. They have been developed on various platforms and used different formats or standards. Todays most widely used image formats include: GIF (Graphics Interchange Format) JPEG (Joi nt Photographic Experts Group standard for compressing still images )2 Animated (Graphics Interchange Format - Animated) BMP (Windows Bitmap format) EPS (Encapsulated PostScript file format) PNG (Portable Network Graphics) PSD (Adobe Photoshops native file format)

We often use JPEG to mean the standard by this expert group, rather than the organisation itself.

Chapter 9: Image data

PSD Layered (Adobe Photoshops native file format) STN (MediaBins proprietary STiNG file format) TIFF (Tagged Image File Format) TGA (TARGA bitmap graphics file format)

Obser vation
From the above discussion, we can see that: 1. Image files are usually large because an image is two-dimensional and can be displayed in so many colours. In a bitmap image, each pixel requires typically 24 bits to represe nt its colour. 2. The loss of some image features is totally acceptable as long as the Human visual system can tolerate. After all, an image exists only for people to view. It is this fact this makes lossy compression possible. 3. Each type of image contains a certain amou nt of redundancy but the cause of the redundancy varies, and this leads to different compression methods. That is why there have to be many different compression meth ods.

Learning outcomes
On completion of your studies in this chapter, you should be able to: illustrate how a picture can be represe nted by either a bitmap image or a vector image explain the main differences between bitmap and vector graphics in terms of visual characteristics illustrate RGB colour model and LC represe ntation describe a few of the commonly used image formats.

Data Compression

Activities
1. If you consider the three components of an RGB colour to be Cartesian coordinates in a 3-dimensional space, and normalise them to lie between 0 and 1, you visualise RGB colour space as a unit cube, see Figure 9.1. (a) What colours correspond to the eight corners of this cube? (b) What does the straig ht line running from the origin to (1,1,1) shown in the figure represe nt? (c) Comme nt on the usefulness of this represe ntation as a means of visualising colour.

(1,1,1) (0,0,0)

B
Figure 9.1: The RGB colour space

2. Classify and identify the following images for a continuous-tone image, discrete-tone image and cartoon-like image: a a a a reproduction of your national flag photograph of the surface of Mars photograph of yourself on the beach still from an old black and white movie.

3. Explain how RGB colour values, when R = G = B, can be used to represe nt shades of grey.

La boratory
1. Investigate the colour utilities (or system, package, software) on your computer. For example, if you have access to Linux, you can go to Start Graphics More Graphics KColorChoser and play with the colour displ ay. 2. Experime nting on how the values of (R,G,B) and (H,S,V) affect the colour displ ayed.

Sample examination

questions

1. Illustrate the RGB colour space using Cartesian coordinates in a 3-dimensional space. 2. Consider a square image to be displayed on screen. Suppose the image dimension is 20 20 square inches. The device resolution is 800 dpi. What would be the pixelDimension required? Solution 1600 1600

Chapter 10: Image compression

Chapter

Image compression
Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 10.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 3.3-3.5, 3.7. Image compression can be lossless or lossy, although most exist image compression algorithms are lossy. We shall introduce the general approa ches in both classes for each type of image data discussed in the previous chapter. However, one should bear in mind that: 1. In general, text compression is lossless by nature: there is no algorithm available to decide independently what text can be omitted. 2. In contrast, many image compression algorithms can delete a lot of image data effectively. 3. The successful performance of the lossy algorithms depends essentially on peoples visual tolerance. First we look at lossless image compression.

Lossless image compression

Lossless approaches mainly extend the compression techniques for text data to image compression. Bi-le vel image Examples include the fax standards: the ITU-T T.4 and T.6 recommendations. Two popular approaches, run-length coding and extended approach are considered below: Run-length coding This approach (see previous chapter) can be applied to bi-level images due to the fact that: 1. each pixel in a bi-level image is represe nted by 1 bit with 2 states, say black (B) and white (W) 2. and the immediate neighbours of a pixel tend to be in an identical colour. Therefore, the image can be scanned in row by row and the length of runs of black and white pixels can be computed. The lengths are then encoded by 81

Data Compression

variable-size codes and are written on the compressed file. Example 10.1 A character A represented by black-white pixels:

12345678901234567890 12345 1 BBBBBBBBBBBBBBBBBBBB BBBBB 2 BBBBBBBBBBBwwwBBBBB BBBBBB 3 BBBBBBBBBBwwBwwBBBB BBBBBB 4 BBBBBBBBBwwBBBwwBBB BBBBBB 5 BBBBBBBBwwBBBBBwwBB BBBBBB 6 BBBBBBBwwBBBBBBBwwB BBBBBB 7 BBBBBBwwwwwwwwwww wwBBBBBB 8 BBBBBwwBBBBBBBBBBBw wBBBBB 9 BBBBwwBBBBBBBBBBBBB wwBBBB 0 BBBBBBBBBBBBBBBBBBBB BBBBB Solution We notice that Line 4 consists of several runs of pixels: 9B, 2w, 3B, 2w, 9B, i.e. 9 Bs followed by 2 ws, followed by 3 Bs, then 2 ws and then 9 Bs. Similarly, line 5 consists of 8B, 2w, 5B, 2w, 8B; ... and so on. A better compression may be achieved by taking into consideration the correlation of adjace nt pairs of lines. For example, a simple idea is to code Line 5 with respect to Line 4 by the differences in number of pixels: -1, 0, 2, 0, -1. Imagine if we send, via a telecom munication channel, the data -1, 0, 2, 0, -1 for Line 5 after the data for Line 4. There would be no confusion nor need to send 7B, 2w, 5B, 2w, 7B which includes more characters. Note: The algorithm assumes that successive lines have the same number of runs. The actual ITU-T T.4 Group algorithm is much more complicated than this. Extended approa ch The idea of this approach extends the image compression principles and concludes that if the curre nt pixel has colour B (or W) then black (or white) pixels seen in the past (or those that will be found in future) tend to have the

same immediate neighbours. The approach checks n of the near neighbours of the curre nt pixel and assigns the neighbours an n-bit number. This number is called the context of the pixel. In principle there can be 2n contexts, but the expected distribution is non-uniform because of the images redundanc y. We then derive a probabili ty distribution of the contexts by counting the occurrence of each context. For each pixel, the encoder can then use adapti ve arithmetic coding to encode the pixel with the probabilities. This approach is actually used by JPEG.

Gr ayscale and colour images

The Graphics Interchange Format (GIF) was introduced by CompuServe in 1987. 82

Chapter 10: Image compression

In GIF, each pixel of the image is an index into a table that specifies a colour map for the entire image. There are only 256 different colours in the whole image. Of course, the colours may also be chosen from a predefined and much larger palette. GIF allows the colour table to be specified for each image, or for a group of images sharing the use of map or without a map at all. Reflected Gr ay codes (RGC) Reflected Gray codes (RGC) is a good represe ntation for grayscale images. In this system, we assign codewords in such a way that any two consecuti ve numbers have codewords differing by 1 bit only. Example 10.2 We show below the decimal numbers that are represented by 1 bit, 2 bit and 3 bit RGC accordingly: decimal value: 0 1 2 3 4 5 6 7 1 bit RGC : 0 1 2 bit RGC : 00 01 11 10 3 bit RGC : 000 001 011 010 110 111 101 100 ... etc. A RGC codeword can be derived from a normal binary codeword as follows: Given a decimal number m, its Reflected Gray codeword is m2 XOR shift-1-bit-to-righ t(m2 ) , where m2 represents the binary codeword of m. Example 10.3 Derive a 3-bit reflected Gray codeword for decimal 3.

The 3-bit binary code of 3 is: 011 Shift 011 one bit to the right (and add 0 in front): 001 011 XOR 001 = 010 So the 3-bit reflected Gray codeword for 3 is 010. Dividing a grayscale image Using Reflected Gray Code (RGC), this approach separates a grayscale image into a number (say, n) of bi-level images and apply to each bi-level image a different compression algorithm depends on the characteristics of each bi-level image. The idea is to assume intuitively that two similar adjace nt pixels in the grayscale image are likely to be identical in most of the n bi-level images. By similar in the examples below, we mean that the number of different bits between two codewords, i.e. the Hamming distance of two binary codewords, is small as well as the difference in value. For example, 0000 and 0001 are similar because their value difference is 1 and the number of different bit(s) is also 1. The two codewords are identical in the first three bits. Now we look at an example of separating a grayscale image into n bi-level images.

Data Compression

Example 10.4 Given a grayscale image with 8 shades of grey, we can represent each shade by 3 bits. Let each of the three bits, from left to right, be identified as the high, middle and low (bit). Suppose that part of a grayscale image is described by matrix A below, where each RGC codeword represents a pixel with the shade of that value: 010 010 011 11 A= 011 010 11 0 001 000 001 011 1 10 1 Then the image can be separated into 3 bi-level images (also called bitplanes) as follows: 1. Bitplane A.high below consists of all the high bits of A. A.high can be obtained from A by removing, for each entry of A, the two bits other than the high bit. 0 0 0 1 A.high = 0 0 0 1 0 0 0 1 2. A.middle below consists of all the middle bits of A. A.middle can be obtained from A by removing, for each entry of A, the two bits other than the middle bit. 1 1 1 1 A.middle = 0 1 1 1 0 0 1 0 3. A.low below consists of all the low bits of A. A.low can be obtained from A by removing, for each entry of A, the two bits other than the low bit. 0 0 1 0 A.low = 1 1 0 1 0 1 1 1 As we can see, from bitplane A.high to bitplane A.low, there is more and more alteration between 0 and 1. The bitplane A.low in this example features more random distribution of 0s and 1s, meaning less correlation among the pixels. To achieve an effective compression, different compression methods should be applied to the bitplanes. For example, we could apply the Run-length meth od to A.high, and A.middle, and the Huffman coding to A.low . Predicti ve enc oding The main idea is to predict the values of a pixel based on some previous pixels and to store the difference between the predicted and actual values in a residual image. Example 10.5 Suppose that the following matrix A[i,j] represents the pixel values of part of a larger grayscale image, where row indices i = 1, , 8 and column indices j = 1, , 8: 84

Chapter 10: Image compression

2 1 3 6 3 3 1 1

4 2 3 1 1 1 1 3 4 4 2 3 2 1 2 5 5 7 4 1 1 3 5 8 3 3 2 2 5 5 7 2 4 3 2 1 5 9 7 4 3 3 1 5 4 9 2 3 3 1 5 5 4 2 3 3

Let us predict that each pixel is the same as the one to its left. So the residual matrix R[i,j]=A[i,j]-A[i,j-1] is as below, where i = 1, , 8, j = 2, , 8: 2 1 3 6 3 3 1 1
i.e. those text comp ression methods for generating a prefix code such as Huffman coding.
1

2 2 1 -2 -1 0 1 0

-2 -1 1 1 5 1 -1 0

1 1 0 3 -4 -1 1 1

-2 2 0 -1 -1 1 1 0

0 -4 0 2 5 0 0 1

0 0 0 -5 2 -2 0 0

0 0 0 1 -5 0 0 0

We can now apply any methods of prefix coding1 , to encode according to the frequency distribution of R. The frequency distribution for this part of the residual image is then as below: entry occurrence probability -----------------------------5 2 2/6 4 -4 2 2/6 4 -2 4 4/6 4 -1 6 6/6 4 0 21 21/6 4 1 16 16/6 46/6 2 6 4 3 4 4/6 4 5 2 2/6 4 6 1 1/6 4 P The entropy of the distribution is i pi log2 pi bits. We can then use this entropy of partial image to code the entire image if the entropy is a good estimation for the whole image. Hopefully this would achieve a better compression in comparison to the original 8 bits/pixel (e.g. ASCII) coding. JPEG lossless coding The predictive encoding is sometimes called JPEG lossless coding because the JPEG standard (codified as the ITU-T T.82 recommendation) specifies a similar predictive lossless algorithm. The algorithm works with 7 possible predictors (i.e. the prediction schemes). The compression algorithm chooses a predictor which maximises the compression ratio. Given a pixel pattern (see below), the algorithm predicts the pixel x? in one of 8 ways: 85

Data Compression

T S Q x?
2

i.e. using the original pixel value matrix (A in Example 10.5), but this tends to be ineff ective.

1. No predictio n2 2. x = Q 3. x = S 4. x = T 5. x = Q + S T 6. x = Q + (S T )/2 7. x = S + (Q T )/2 8. x = (Q + S)/2 The lossless JPEG3 algorithm gives appr oximately 2:1 compression ratios on typical grayscale and colour images, which is well superior to GIF (although GIF may be competitive on icons or the like).

It is generally for lossy standard but does have the lossless version and It is an extensive and complicated standard for common use.

Lossy compression
Lossy compression aims at achieving a good compression ratio, but the cost for it is the loss of some original information. Therefore, the measure of a lossy compression algorithm normally includes a measure of the quality of reconstructed images compared with the original ones. Distortion measure Although the best measure of the closeness or fidelity of a reconstructed image is to ask a person familiar with the work to actually look at the image and provide an opinion, this is not always practical because it is not useful in mathematical design approa ches. Here we introduce the more usual approach in which we try to measure the difference between the reconstructed image and the original one. There are mathematical tools that measure the distortion in value of two variables, also called difference distortion measure. Considering an image to be a matrix of values; the measure of lossy compression algorithm normally uses a standard matrix to measure the difference between a reconstructed images and the original ones. Let Pi be the pixels of reconstructed image and Qi be the ones of the original, where i = 1, , N . We have the following commonly used measures: Squared error measure matrix (this is a measure of the difference): Di = (Pi Qi )
2

Absolute difference measure matrix: Di = |Pi Qi |

Chapter 10: Image compression

Mean squared error measure (MSE) matrix (this is an average measure): 1 X i Qi )2 =N (P i= 1

2 N

Signal-to-noise-ratio (SNR) matrix (this is the ratio of the average squared value of the source output and the MSE): SNR = This is often in logarithmic scale (dB): SNR = 10 log10
2 x 2 d 2 x 2 d

Peak-signal-to-noise-ratio (PSNR) matrix (this measures the error relative to the average squared value of the signal, again normally in logarithmic scale (dB)): SNR = 20 log10 max i |Pi | 2 d

Average of the absolute difference matrix: Di =

N 1 X |Pi Qi | N i

Maximum value of the error magnitude matrix: Di = max Di = |Pi Qi |

Progressi ve image compression The main idea of progressive image compression is to gradually compress an image following a underlined order of priority. For example, compress the most importa nt image information first, then compress the next most importa nt information and append it to the compressed file, and so on. This is an attracti ve choice when the compressed images are transmitted over a communication channel, and are decompressed and viewed in real time. The receiver would be able to view a development process of the image on the screen from a low to a high quality. The person can usually recognise most of the image features on completion of only 5-10% of the decompression. The main advantages of progressive image compression are that: 1. The user can control the amou nt of loss by means of telling the encoder when to stop the encoding process 2. The viewer can stop the decoding process early since she or he can recognise the Images feature at the early stage 3. as the compressed file has to be decompressed several times and displayed with different resolution, the decoder can, in each case, stop the decompression process once the device resolution has been reached. There are several ways to impleme nt the idea of progressive image compression: 87

Data Compression

1. Using so-called SNR progressive or quality progressive compression (i.e. encode spatial frequency data progressi vely). 2. Compress the gray image first and then add the colour. Such a meth od normally features slow encoding and fast decoding. 3. Encode the image in layers. Early layers are large low-resolution pixels followed by later smaller high-resolution pixels. The progressi ve compression done in this way is also called pyramid coding or hierarchical coding. Example 10.6 The following methods are often used in JPEG:

Sequential coding (baseline encoding): this is a way to send data units following a left-to-right, top-to- bottom fashion. Progressive encoding: this transmits every nth line before filling the data in the middle. Hierarchical encoding: this is a way to compress the image at several different resolutions. Transforms Transform is a standard mathematical tool used to solve complicated computation problems in many areas. The main idea is to change a quantity such as a number, vector or a function to another form in which some useful features may occur (e.g. the computation in the new form may be easier, or becomes more feasible). The result is then transformed back to the original form. An image can be compressed if its correlated pixels are transformed to a new represe ntation where the pixels become less correlated, i.e. are decorrelated. Compression is successful if the new values are smaller than the original ones on average. Lossy compression can be achieved by quantisation of the transformed values. The decoder normally reconstructs the original data from the compressed file applying the opposite transform. Orthogonal transform In this transform, matrix computation is applied. Let C and D be vectors and W be a n n matrix and ci , di and wij be the elements in C , D and W P respectively. Further let ci = j wij dj , where ci is a weighted sum of pixels dj of an image and each pixel is multiplied by a weight wij , i, j = 1, , 3. This can be written in matrix form: w11 c1 c2 = w21 w31 c3 w12 w22 w32 w13 w23 w33 d1 d2 d3

We can write C = W D. Each row of W is called basis vector. In practice, the weights should be independent of the values of di s so the weights do not have to be included in the compressed file.

Chapter 10: Image compression

Example

10.7 Given a weight matrix: 1 1 1 W = 1 1 1 1 1 1 4 D = 6 5

If we use the transfer vector

Then we have

1 1 1 4 15 1 1 1 6 = 7 1 1 1 5 3

Compare the energy of (15, 7, 3)T to that of (4, 6, 5)T . We have 152 + (7)2 + 32 = 283 compared with 42 + 62 + 52 = 77, which is nearly 4 times smaller. However, the original energy may be conserved if we adjust W by multiplying a factor of 1/2. Now we have C = (15/2, 7/2, 3/2) T . When reconstructing the original D, we first quantify the transformed vector C from (7.5, 3.5, 1.5)T to integers (8, 4, 1)T and conduct an inverse transform: D = W1 C 1 1 1 8 3 (1/2) 1 1 1 4 = 5 1 1 1 2 7 In practice, the W is much bigger so the reconstruction result would be better. More sophisticated transform techniques would also produce better results. Wavelet transform s Wavelet transform is one of the most popular forms of sound or video compressing signals. The two-dimensional Wavelet transform can be used for image compression. The main idea is similar to an extended version of the one discussed in the previous section. Wavelet transform can require a fair degree of mathematical sophistication to impleme nt and has many tool kits available including an entire class of special processing chips and devices, known as digital signal processors. Due to time pressure, we only briefly sketch a general picture without details in this subject. Also, for convenience of discussion, we give the definition of one-dimensional transform only. The Wavelet transform works by choosing a set of represe ntative patterns and finding a subset of these patterns that add up to the signal. Let f (t) be the signal and t is the measure of time. This might be a sound file that stores the intensity of the sound every millisecond. Let gi (t) be another set of functions known as the basis. Suppose a function f can be represe nted as the weighted sum of some set of gi (t):
N X f (t) i gi (t) i

Data Compression

The values of i are called the coefficients, which can be boolean, integer or real numbers. The compression process aims to find a good set of coefficients (1 , 2 , , N ) to represe nt signal f . The decompression process aims to reconstruct f from the set of coefficients. The compression and decompression algorithms share the same set of functions and the compressed file consists of the set of coefficients. The main issues in a Wavelet transform are: finding a good set of basis functions for a particular class of signal finding an effective quantisation method for different frequencies. At present, the effort and choices made for both of the issues are as much as an art rather than a science. Due to time limits, we only list the most common approaches here: Discrete Cosine (Sine) transform (DCT) Fourier transform (FT) Walsh-Hadamard Transform (WHT) Haar Transform (HT)

Kar hunen-L oeve Transform (KLT) Interested readers may consult the books given at the beginning of the chapter for details.

JPEG (Still) Image Compression

Standard

This is one of the most widely recognised standards in existence today. and provides a good example of how various techniques can be combined to produce fairly dramatic compression results. The baseline JPEG compression method has a wide variety of hardware and software impleme ntations available for many applications. JPEG has several lossy encoding modes, from so-called baseline sequential mode to lossless encoding mode. The basic steps of lossy JPEG algorithm include processing 24 (or 32) bit colour images and offering a trade-off between compression ratio and the quali ty. There are a lot of techniques and algorithms for image data compression. You may consult various sources in literature. Due to the time limit in the subject, we have only provided a sample of the techniques here.

Learning outcomes
On completion of your studies in this chapter, you should be able to: provide examples of lossless and lossy image compression techniques illustrate how to represe nt a grayscale image by several bi-level images describe the main ideas of predictive encoding be familiar with various distortion measureme nts explain the general approaches of various popular lossy image compression techniques such as progressive image compression and transforms. 90

Chapter 10: Image compression

Activities
1. Construct a Reflected Gray Code for decimal numbers 0, 1, , 15. 2. Following the above question, provide an example to show how a grayscale image with 16 shades of gray can be separated into 4 bi-level images. 3. Suppose that the matrix A[i, j] below represents the pixel values of part of a large grayscale image, where i = 0..7, j = 0..7, and A[0, 0] = 4. Let us predict that each pixel is the same as the one to its right. (a) Derive the residual matrix R[i, j] of A (b) Derive the frequency table for R (c) Derive a variable length code, e.g. Huffman code for the partial image. (d) Discuss under what conditions the code can be used for the whole image. 4 1 8 2 2 2 3 7 8 2 4 4 4 2 3 7 4 4 5 8 6 2 6 7 8 6 5 5 7 3 6 7 1 5 5 7 7 4 6 6 1 1 5 9 7 9 7 7 1 1 5 5 9 7 7 8 1 1 5 5 9 3 7 8 1 1 1 1 1 1 1 1 W = 1 1 1 1 1 1 1 1 4 6 , D = 5 2

4. Given the orthogonal matrix

and a data vector

conduct a similar experime nt to the approach in this chapter. (a) Fulfil a linear transform on D (b) Perform a reverse transform to get D 0 (c) Compare D0 with the original D. 5. Impleme nt the ideas above using a higher level programming language such as Java or C.

La boratory
1. Design and impleme nt a method in Java which takes a binary codeword and returns its RGC codeword. 2. Design and impleme nt a method in Java which takes a matrix of integers and returns its residual matrix, assuming that each entry is the same as the one to its left. 91

Data Compression

Sample examination

questions

1. Derive the RGC codeword for a decimal 5. 2. Explain, with an example, how to represe nt the part of a grayscale image below by 3 bitplanes: 00 00 0 1 01 1 00 00 1 1 00 1 01 00 1 1 01 0 01 01 1 0 00 0

Chapter 11: Video compression

Chapter

Video compression
Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 11.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 5.

Video systems
These can be classified into two types: analog video and digital video. Analog video The essential function of an analog video is to display still pictures on a television receiver or a monitor screen one frame after another. The pictures are converted into an electronic signal by a raster scanning process. Conceptuall y, a screen is divided into horizontal lines. The picture we see is built up as a sequence of horizontal lines from top to bottom. Here the human persistence of vision plays an importa nt part and makes the series of lines appear a (frame of ) picture. The following parameters determine the quality of pictures: Number of scanning lines Number of pixels per scan line Number of displayed frames per second (in fps): this is called the frame rate Scanning techniques. For example, the screen has to be refreshed about 40 times a second, it flickers otherwise. Popular standards include Phase Alternating Line (PAL), Seque ntial Couleur Avec Memoire (SECAM) and (USA) National Television Systems Committee (NTSC). PAL is mainly used in most of Western Euro pe and in Australia, China and New Zealand. SECAM is used in France, in the former Soviet Union and in Eastern Europe. NTSC is used in North America, Japan, Taiwan, part of the Cribbean, and in South America. A PAL or SECAM frame contains 625 lines, of which 576 are for pictures. An NTSC frame contains 525 lines, of which 480 are pictures. PAL or SECAM use a frame rate of 25 fps and NTSC 30 fps. Digital video Although the idea of digital video is simply to sample the analog signal and convert it into a digital form, the standard situation is inevitably quite complex due to the need to compromise the existing equipme nt in the past as well as new standards such as HDTV. The sample standard is ITU-R BT.601 (more commonly known as CCIR 93

Data Compression

601), which defines sampling of digital video. It specifies a horizo ntal sampling picture format consisting of 720 luminance samples and two sets of 360 colour difference samples per line. The size of PAL screen frames is 768x576 and NTSC 640x480. Besides the digital data from analog video sampling, digital video data to be compressed also includes the data streams produced by various video equipme nt such as digital video cameras and VTRs. Once the data is input into the computer there are many ways to process them.

Moving pictures
Video data can be considered as a sequence of still images changing with time. Use of human eyes forgiving feature, video signals change the image 20 to 30 times per second. Video compression algorithms all tend to be lossy. One minute of modern video may consist of 1,200 still images. This explains the need and the moti vation for video compression. Among different numbers of different video compression standards, two are the most importa nt ones, namely ITU-T H.261 and MPEG. ITU-T H.261 is intended for use mainly for video conferencing and video-telephony, and MPEG is mainly for computer applications.

M PEG
This is the standard designed by the Motion Picture Expert Group, including the updated versions MPEG-1, MPEG-2 MPEG-4, MPEG-7. Most popular standards include the ISOs MPEG-1 and the ITUs H.261. It is built upon the three basic common analog television standards: NTSC, PAL and SECAM There are two sizes of SIF (Source Input Format): SIF-525 (with NTSC video) and SIF-625 (for PAL video). The H.261 standard uses CIF (the Common Intermediate Format) and the QCIF (Quarter Common Intermediate Format). As we can see, MPEG is greatly influenced by ITU-T standards and even includes some of its standards. This may be the reason that ITU-T H.261 is the more fully developed of the two.

Basic principles
Video is based on two characteristics of the video data:
Recall this means the correlation among neighbouring pixels in an image.
2 1

spatial redundanc y1 exists in each frame: this can be dealt with by the techniques for compressing still images. tem poral redundanc y2 occurs because a video frame tends to be similar to its immediate neighbours. Techniques for removing spatial redundancy are called spatial compression or infra-frame and techniques for removing temporal redundancy are called temporal compression or inter-f rame. We focus on temporal compression here. Temporal compression algorithms In these algorithms, certain frames in a sequence are identified as key frames . These key frames are often specified to occur at regular intervals. The key

Recall this means the similarity among neighbouring frames.

Chapter 11: Video compression

frames are either left uncompressed or are more likely to be spatially compressed. Each of the frames between the key frames is replaced by a so-called difference frame which records the differences between the original frames and the most recent key frame. Alternati vely, the differences between the original frames and the preceding frame can also be stored in the difference frame depending on the sophistication of the decompressor. Example 11.1 Figure 11.1 shows a sequence of frames to be displayed during a time duration 0, 1, , 11.

1 time: 0 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 10 11 9 10 11

Figure 11.1: A sequence of original frames Follows the MPEGs regulation, the frames marked I (as shown in Figure 11.2) are encoded as the key frames. These are called I-pictures, meaning intra and having been compressed purely by spatial compression. Difference frames compared with previous frames are called P-pictures or predictive pictures. The frames marked B are so-called B-pictures which are predicted from later frames. Example 11.2 The encoded frames are as Figure 11.2.

1 time: 0 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 10 11 9 10 11

Figure 11.2: Encoded as I-,P- and B pictures The decoding process now actually becomes a question of dealing with a sequence of compressed frames with I,P,B types of pictures. I-pictures are decoded independently. P-pictures are decoded using the preceding I or P frames. B-pictures are decoded according to both preceding and following I- or P-pictures. The encoded frames can then be decoded as IBBBPBBBP and are displayed in that order (see Figure 11.4). Example 11.3 Note that the order of the encoded and displayed frames in Figure 11.3 are altered slightly from those in Figure 11.2.

Data Compression

1 time: 0 1 2

3 3

4 4

5 5

2 6

8 7

9 8 9

6 11 10 10 11

display

1 time: 0 1 2

3 3

4 4

5 5

2 6

8 7

9 8 9

6 11 10 10 11

Figure 11.3: Decoded and displayed from I-,P- and B pictures Group of pictures (G P) O Now a video clip can be encoded as a sequence of I-,P- and B-pictures. GOP is a repeating sequence beginning by an I-picture and is used by encoders. Example groups. 11.4 Figure 11.4 shows a GOP sequence IBBPBB containing 2

B B

time: 0

10 11

Figure 11.4:

Other common sequences in use are IBBPBBPBB with 3 groups and IBBPBBPBBPBB with 4 groups.

Motion estimation
The main idea is to represe nt each block of pixels by a vector showing the amou nt of block that has moved between images. The approach is most successful for compressing a background of images where the fixed scenery can be represe nted by a fixed vector of (0,0). It uses the following meth ods: Forward prediction - computing the difference only from a previous frame that was already decoded. 96

Chapter 11: Video compression

Backward prediction - computing the difference only from the next frame that is about to be decoded. Bidirectional prediction - the pixels information in a frame depends on both the last frame just decoded and the frame about to follow. The following factors can be used to estimate motion: Basic difference. L1 norm - mean absolute error ||A B|| 1 = L2 norm - mean squared error ||A B|| 2 = X X
i j

X X
i j

|Ai,j Bi,j |.

(Ai,j Bi,j )2 .

Work in differe nt video formats

To allow systems to use different video formats, a process as follows is needed: Standardise the Format. Break the data into 8 8 blocks of Pixels. Compare to other Blocks. Choose importa nt coefficients. Scale the coefficients for quantisation. Pack the coefficients.

Learning outcomes
On completion of your studies of this chapter, you should be able to: identify and list the parameters that determine the quality of pictures explain the popular standards for analog and digital videos describe, with an example, how temporal compression algorithms work in principle explain the common measures for motion estimation.

Data Compression

Activities
1. Explain briefly the concepts of spatial redundancy and temporal redundanc y. 2. Impleme nt the basic temporal compression algorithm introduced in the chapter.

Sample examination

questions

1. Explain the two characters of video data. 2. Illustrate how a temporal compression algorithm works.

Chapter 12: Audio compression

Chapter

Audio compression
Esse ntial reading
Wayner, Peter Compression Algorithms for Real Programmers . (Morgan Kaufmann, 2000) [ISBN 0-12-788774-1]. Chapter 12.

Further

reading

Salomon, David A Guide to Data Compression Methods. (Springer, 2001) [ISBN 0-387-95260-8]. Chapter 6.

Sound
Sound is essentially a sensation of the human audio system (i.e. it is detected by the ears and interpreted by the brain in a certain way). It is similar to colour in a way, a mixture of physical and psychological factors and it is therefore not easy to model accuratel y. Nevertheless, we know that sound is a physical disturbance in a medium and it is propagated in the medium as a pressure wave by the movement of atoms or molecules. So sound can be described as a function (as for electronic signals), f (t) measuring the pressure of medium at a time t. The easy function for this purpose is a sine wave p(t) = sin (t) = sin(2f t). The f is the frequency describing the number of cycles per second and it is measured in Hertz (Hz). If T is the period, the sine wave can be written as p(t) = p(t + T) and sin (t) = sin( T2 t) since f =T 1 . Any periodic signal s(t) with period T can be represe nted by the sum of sine or cosine waves: a0 2
X i=1

s(t) =

a cos(i i T

t) +

X i

i=1

a sin(i T

Example 12.1 We know that most orchestras tune the A above middle C note to 440Hz. The sound pressure function for that note can be written therefore as p(t) = sin(880t). We also know that the human ear is normally able to detect frequencies in the range between 20Hz and 20kHz. The upper audible limit in terms of frequency tends to decrease as age increases. Children may hear sounds with frequency as high as 20kHz but few adults can. It is interesting to notice that the timbre of different instrume nts is created by the transie nt behaviour of notes. In fact, if the attack portion is removed from recordings of an oboe, violin, and soprano playing or singing the same note, the steady portions are indistinguishable. Like any other wave functions, we can plot any sound by plotting its amplitude (pressure on medium) against time. The shape of the curve can demonstrate clearly the different properties of a specific sound. 99

Data Compression

Alternati vely, we can use a frequency spectrum diagram to represe nt the changes of frequencies over a period of time. In a frequency spectrum diagram, we plot the amplitude against the frequenc y. Example 12.2 Figure 12.1 is a frequency spectrum diagram for s(t) = 0.1 sin( 2t) + 0.3 sin( 8t) + 0.25 sin(20t).

Amplitude 0.3 0.2 0.1 0 1 2 3 4 5 6 7 8 9 10 Frequency (Hz)

Figure 12.1: Frequency spectrum diagram

Digital sound data

Digitised audio data are the digital form of sounds. They are created by a process of sampling followed by quantisation on the analog sound wave. Sampling Sampling is a way of taking certain values at n discrete time t1 , t2 , , tn . Samples taken at a time series with an equal time interval is called sampling at a fixed rate. The sampling frequency rate is simply 1/t . Otherwise, sampling taken with various different time intervals are called sampling at a variable rate. In this case the sampling frequency can be measured by the number of samples taken per second. Example 12.3 The sampling rate used for audio CDs is 44.1 kHz, 48 kHz is used for DAT (digital audio tape). 22.05 kHz is commonly used for audio signal for delivery over the Internet, and 11.025 kHz may be used for speech. Nyquist frequency It is not difficult to understand that the choice of sample frequency rate can directly affect the quality of the digital sound data and subseque ntly the quali ty of compression. According to Nyquist theory, if a continuous wave form is band-limited to a frequency f then it must be sampled at a frequency of at least 2f in order to be able to reproduce the original waveform exactl y. Digitisation Usually the values achieved from sampling are real numbers. Specifically, we need to convert the values into integers within the range [0,k-1] for some suitable integer k. For sampling or digitisation, we may use the techniques introduced in previous chapters, or other more efficient techniques.

100

Chapter 12: Audio compression

Audio data
If we plot the amplitude over a period of time for both the sound generated by speech and by a piece of music, we would find that the shape of the two waveforms very different. This suggests that we should deal with them separatel y. Two types of sounds with very different characteristics are our main consideration here: speech and music. They are the most commonly used types of sound in multimedia productions today. Represe ntations specific to music and speech have been developed to effectively suit their unique characteristics. Speeches are represe nted in models of sounds, but music is represe nted as instructions for playing virtual instrume nts. This leads to two compression areas, namely voice compression and music compression. Voice compression aims at removing the silence and music compression at finding an efficient way to reconstruct music at events. 1 In general, audio compression approaches are lossy.

The term event is used to mean any form of music, not merely live social events

Speech compression
This is also called voice compression. We only briefly introduce a few commonly used compression methods here. Pulse code modulation (ADPCM ) This technique is well known in other research areas such as telecom munication and networking. ADPCM stands for Adaptive Differential Pulse Code Modulation. It was formalised as a standard for speech compression in 1984 by ITU-T. The standard specifies compression of 8 bit sound (typical represe ntation for speech) sampled at 8 kHz to achieve a compression ratio of 2:1. Speech coders Here two major types of audio data are considered: telephone speech wideband speech The goal is to achieve a compression ratio of 2:1 or better for both types of data. More specifically, the aim is to compress telephone speech to the compressed bit rate of less than 32 kbps, and to 64 kbps for wideband speech. The following constrai nts make it possible to fulfil some of the tasks: The human ear can only hear certain sounds in normal speech. The sounds produced in normal human speech are limited. Non-speech signals on telephone lines are noises. Two coders are used in practice, namely waveform coders and vocoders (voice coders). They both apply detailed models of the human voice tract to identify certain types of patterns in human speech. 101

Data Compression

Predicti ve approa ches This is similar to our approaches for video compression. The idea is to try to predict the next sample based on the previous sample and code the differences between the predicted and the actual sample values. A basic version of the impleme ntation would simply apply entropy coding, such as Huffman or arithmetic, to the differences between successive samples. This approach is used in so-called VOCPACK algorithm for compressing 8 bit audio files (.wav). Example 12.4 Given a series of sample data, (27,29,28,28,26,27,28,28,26,25,27). We can write the difference (2,-1,0,-2,1,1,0,-2,-1,2). Now the entropy of the distribution of differences is log2 5 = 2.32 bits. This would give a compression ratio of 8:2.23 (i.e. 3.45:1 if this can apply to the entire file). Of course, sampling is also a critical step to achieve good compression and it is always worth considering.

Music compression
Compression algorithms for music data are less well developed than that for other types of data. The following facts contribute to this situation: There are various formats for digital sounds on computer. The development of digital audio has taken place in recording and broadcast industries. There are currently three major sound file formats: AIFF of MacOS, WAV or WAVE for Windows and AU (NeXT/Sun audio file format). Each of the three have evolved over the years and now provide similar capabilities in terms of the sampling rates and sizes and CD and DAT standard value storage, while MP3 in its own file format cannot accomm odate sound compression by any other meth od. We usually use a high bit rate and often attempt to capture a wider range of frequencies (20 to 20k Hz). It is usually difficult to decide what is a good compression system because this depends on each persons hearing abili ty. For example, to most people, lower frequencies add bass resonance and the result would sound more like the human voice. This, however, may not be the case at all for others. Streaming audio The idea of streaming audio is to deliver sound over a network and play it as it arrives without having to be stored on the users computer first. This approach is more successful in general for sound than it is for video due to the lower requireme nt for bandwidth. The available software includes: Real Networks RealAudio (companion to RealVideo) 102

Chapter 12: Audio compression

Streaming QuickTime lo-fi MP3 These are used already for broadcasting live concerts, for the Internet equivalent of radio stations, and for providing a way of playing music on the Internet. M IDI The main idea of MIDI compression is, instead of sending a sound file, to send a set of instructions for producing the sound. Of course, we need to make sufficiently good assumptions about the abilities of the receiver to make sure the receiver can play back the sound following the instruction, otherwise the idea cannot work in practice. MIDI stands for Musical Instrume nts Digital Interface. It was originally a standard protocol for communicating between electronic instrume nts. It allowed instrume nts to be controlled automatically by sound devices that could be programmed to send out MIDI instructions. MIDI can control many sorts of instrume nts, such as synthesisers and samplers, to produce various sounds of real music instrume nts. Examples of available software include: QuickTime Cakewarlk Metro Cubse QuickTime incor porates MIDI-like functionali ty. It has a set of instrume nt samples and can incor porate a superset of the features of MIDI. It can also read standard MIDI files, so any computer with it installed can play MIDI music without any extra requireme nt. Note: One should realise that sound tends to work together with pictures or animation. In future most audio work has to take any potential synchronisation into consideration. For example, sound divided into video frames by some time code would be a useful function for film editing.

Learning outcomes
On completion of your studies in this chapter, you should be able to: explain how sound can be represe nted by a periodic function illustrate sound by frequency spectrum diagram as well as the normal plot for periodic functions describe the concept or principle of terms in audio data compression, such as ADPCM outline the distinction between voice compression and music compression in terms of the issues concerned in audio compression describe the main ideas of MIDI compression.

103

Data Compression

Activities
1. Given that singing has characteristics of both speech and music, which compression algorithms would you expect to be most successful on songs? 2. Given a frequency spectrum diagram as in Figure 12.2, write the signal s(t) in its analog form, i.e. represe nted by the sum of sine and cosine waves.

Amplitude 0.3 0.2 0.1 0 1 2 3 4 5 6 7 8 9 10 Frequency (Hz)

Figure 12.2:

3. Explain what MIDI stands for. Write a short essay of about 500-1000 words to introduce any application software which uses MIDI in one way or another.

Sample examination
(a) Sampling (b) ADPCM (c) MIDI

questions

1. Explain briefly the following terms:

2. Explain how a frequency spectrum diagram can be used to represe nt function f (t) = 0.5 + sin(880 t) + sin (1760t).

104

Chapter 13: Revision

Chapter

Revision
Revision materials
You should make a good use of the following: lecture notes subject guide lab exercises exercises courseworks text book(s) any other handouts

Other references
You are advised to review mainly the revision materials listed above. If time permits, you may also like to look up some concepts or different versions of algorithms introduced in lectures from various books. However, this is not essential for the examination pur pose.

Examination
The duration of the examination will be two hours and 15 minute s.
1

The three-out-of-five pattern starts from 2004

There will be five questions and you must answer three out of the five question s1 . There will be 75 marks available on the examination paper, but your marks will be converted to a right percentage later, of course. There will be a number of sub-questions in each question.

Questions in examination
The questions in the examination can be classified into the following three types: Bookwork The answers to these questions can be found in the lecture notes listed above. Similar Question The questions are similar to examples in the revision materials listed above. These are usually questions that require examples to be provided in addition to the definitions and explaining some concepts. Unseen Question You may have not seen these types of questions before the exam, but you should be able to answer these questions using the knowledge and experience gained from the course. The questions require some deeper understanding of the knowledge from the stude nts. The questions can also be classified according to the level of requireme nts to the stude nt. For example, Definition question the answer to these questions require merely basic knowledge or definition in lecture notes. Questions are likely to 105

Data Compression

begin with What is xxx .. ?, Explain what xxx is ... Conceptual question This type of questions require from the stude nt a good understanding of some importa nt concepts. The questions seek mainly some concepts to be reviewed over what you have learnt. The questions are likely to start with Discuss xxx .. ?, Explain why xxx ... The answers to these questions should include not only the definition, theorem etc. learnt from the course, but also supporting examples to demonstrate good understanding of issues related to the concepts. Do-it- yourself question These questions require deeper understanding than that for Conceptual questions. The answers to the questions require not only the knowledge gained from the course but also certain degree of problem-solving skills involved. The questions normally consist of instructions to complete certain tasks. For example, Write a piece of Java code to ..., Draw a diagram for ..., Traverse the graph ..., or Analyse xxx ... etc.

Read questions
Every year stude nts are advised to read the questions on the examination paper carefully. You should make sure that you fully understand what is required and what subsections are involved in a question. You are encouraged to take notes, if necessary, while attempting the questions. Above all, you should be completely familiar with the course materials. To achieve a good grade, you need to have prepared well for the examination and to be able to solve problems by applying the knowledge gained from your studies of the course.

Recommendation
The following activities are recommended in your revision: 1. Write the definitions and new concepts learnt from the course in your own words. 2. Find a good example to support or help explaining each definition or concept discussed in the class. 3. Review the examples given in the Revision materials. Try to solve some problems using the skills learnt from the course.

Importa nt topics
These will be covered in the examination. 1. Introduction Entropy Prefix codes Optimal codes Expected length, i.e. average length, of a code Unique decodabili ty 2. Minimum redundancy coding Run-length algorithms Shannon- Fano coding

Huffman coding Adaptive Huffman coding 106

Chapter 13: Revision

3. Arithmetic coding Algorithm Probabili ty trees 4. Dictionary based compression LZW compression Sliding window (LZ77) compression (knowledge) 5. Image lossless compression Image data Colour systems 6. Concepts and principles of Lossy compression Predicti ve coding 7. Speech compression (knowledge)

Goo d luck!
I hope you find this subject guide stimulating and useful, and you have enjoyed the studies as much as I do. Please feel free to let me know ([email protected]) how you feel about the whole experience of studying this module, the good, the bad or both. Finally, I wish you all the best in the exam!

107

Data Compression

108

Appendix A: Sample examination paper I

Ap pendix A

Sample examination
Data Com pression
Duration: 2 hours and 15 minutes

pa per I

You must answer THREE questions only. The full mark for each question is 25. The marks for each subquestion are displayed in brackets (e.g. [3]). Electronic calculators may be used. The make and model should be specified on the script and the calculator must not be programmed prior to the examination. Question A.1

1. Determine whether the following codes for the alphabet {a, b, c, d} are uniquely decodable. Give your reasons for each case. [8] (a) {1, 011, 000, 010} (b) {1, 10, 101, 0101} (c) {0, 01, 011, 0111} (d) {0, 001, 10, 011} 2. Consider a text model for a black-white source file. Under what probability distribution condition does a static Huffman encoding algorithm achieve the best performance? Under what condition on the source does the algorithm perform the worst? Give your reasons. [7] 3. Outline a simple version of the adaptive Huffman encoding algorithm. [5] 4. Illustrate how adaptive Huffman encoding works on a source of text CAAABB. [5] Hint: (a) You may demonst rate step by step the progress of running the Adaptive Huffman encoding algorithm in terms of the input, output, alphabet and the tree structur e. (b) Add sufficient comments in your algorithm if it is different from the one discussed in lecture.

109

Data Compression

Question

A.2

1. Consider part of a grayscale image with 16 shades of gray that is represented by the array A below: 100 110 100 0 0 100 0 110 0 0 Demonst rate how the (bi-level images) A: 100 001 110 011 0 0 0 100 0 100 1 1 image can be represented by several bitplanes [4]

2. Explain, with the aid of an example, why a Huffman code is in general not optimal unless the probability of every symbol in the source alphabet is a negative power of 2. [5] 3. One way to improve the efficiency of Huffman coding is to maintain two sorted lists during the encoding process. Using this approach, derive a canonical minimum variance Huffman code for the alphabet {A,B,C,D,E, F} with the probabilities (in %) 34,25,13,12,9,7 respectively. [8] 4. Explain, with the aid of an example, each of the following terms: [8] (a) fixed-to-variable model (b) gray-scale image

110

Appendix A: Sample examination paper I

Question

A.3

1. Explain what is used to represent the so-called colour depth in a common RGB colour model. What is the value of the colour depth in a representation where 8 bits are assigned to every pixel? [4] 2. Comment on the truth of the following statement describing the absolute limit on lossless compression. [4] No algorithm can compress even 1% of all files, even by one byte. 3. Explain briefly what is meant by Canonical and Minimum-varian ce Huffman coding, and why it is possible to derive two different Huffman codes for the same probability distribution of an alphabet. [4] 4. Describe briefly, with the aid of an example, how Shannon- Fano encoding differs from static Huffman encoding. [5] Hint: You may give an example by deriving a code for {A,B,C, D} with probability distribution 0.5, 0.3, 0.1, 0.1 using the two algorithms, and show the difference. 5. A binary tree (0-1 tree) can be used to represent a code containing a few codewords of variable length. Consider each of the four codes ((i)(iv) below) for alphabet {A,B,C, D} and draw the binary tree for each code. (a) (b) (c) (d) {000, 001, 110, 111} {110,111,0, 1} {0000,0001,1,00 1} {0001,0000,0001, 1}

For each tree drawn, comment on whether the code being represented by the binary tree is a prefix code and give your reasons. [8]

111

Data Compression

Question

A.4

1. Explain the meaning of the following terms: [8] (a) rate-distortion problem (b) entropy (c) prefix codes (d) motion prediction. 2. Given an alphabet of four symbols {A,B,C, D}, discuss the possibility of a uniquely decodable code in which the codeword for A has length 1, that for B has length 2 and for both C and D have length 3. Justify your conclusions. [4] 3. Derive the output of the HDC algorithm on the source sequence below. Explain the meaning of each control symbol that you use. Finally, describe briefly a data structu re(s) and algorithm(s) that can be used in solving counting subproblems in the HDC algorithm. [5] TTUKRR33333333333333PPPEE Hint: You may simply describe the data structu re(s) and algorithm(s) that you have used in your Lab Exercise when implementing the HDC algorithm. 4. Explain the concept of bitmapped images and vector graphics. What are the differences between vector and bitmapped graphics in terms of requirements to the computer storage capacity, and the size of the image files. [8]

112

Appendix A: Sample examination paper I

Question

A.5

1. It has been suggested that the unique decodability is not a problem for any fixed length code. Explain why this is so with the aid of an example. [4] 2. Describe the main idea of predictive encoding. Suppose the matrix below represents the pixel values (in decimal) of part of a grayscale image. Let the prediction be that each pixel is of the same value as the one to its left. Illustrate step by step how predictive encoding may be applied to the array. [6] 1 111 5 111 5 555 7 945 3. Demonst rate step by step how the Basic LZW encoding and decoding algorithms maintain the same version of a dictionary without ever transmitting it between the compression and the decompression end, using a small source string BBGBGH as an example. The LZW encoding and decoding algorithms are given below.
(a) Encoding 1. w ord = ; 2. while not end_of_file x=read_next_characte r; if word+x is in the dictionary word=word +x else output the dictionary index for word; add word+x to the dictionary; word=x ; 3. output the dictionary number for word; (b) Decoding 1. read a token x from compressed file; 2. look up dictionary for element at x; output element word=eleme nt; 3. while not end_of_compressed_file do read x; look up dictionary for element at x; if there is no entry yet for index x then element=word+first_char_of_word; output element; add word+first_char_of_element to the dictionary; word=eleme nt; 4. end

[15]

113

Data Compression

114

Appendix B: Exam solutions I

Ap pendix B

Exam solutions I
Please be aware that these solutions are given in brief note form for the purpose of this section. Candidates for the exam are usually expected to provide coherent answers in full form to gain full marks. Also, there are other correct solutions which are completely acceptable but which are not included here. Duration: 2 hours and 15 minutes You must answer THREE questions only. The full mark for each question is 25. The marks for each subquestion are displayed in brackets (e.g. [3]). Electronic calculators may be used. The make and model should be specified on the script and the calculator must not be programmed prior to the examination. Note: B for bookwork: The answer to the question can be found in the lecture notes or study guide; S for similar: The question is similar to an example in the lecture notes or study guide; U for unseen: The stude nts may have not seen the question but good students should be able to answer it using their knowledge and experience gained from the course. ..................................................................... ........ Question 1. (S) (a) Yes, because this is a prefix code. [1+1/8 ] A.1

(b) No. Because, for example, 01011010101 can be decoded either as DCD or as DBBC (or DBAD). [1+1/8 ] (c) Yes, because each codeword starts with a 0. [1+1/8 ] (d) No, because for example, 00100010 can be decoded as BAA AC or AACBA . [1+1/8 ] 2. (U) Let the probability distribution of the source be {pB , pW }. When pB = pW = 1/2, the static Huffman algorithms work best. [1/7] [1/7]

The reason is that the average length of the Huffman code is equal to the entropy of the source. The Huffman code is optimal in this case. The average length of the codewords is 1 and the entropy of the source is 1/2( log2 1/2)1/2( log2 1/2)+ = 1. [2/7] 115

Data Compression

When pB = 0, pW = 1 (or pB = 1, pW = 0), the static Huffman algorithms work worst.

[1/7]

The average length of the codewords is 1 and the entropy of the source is 1( log2 1) = 0. [2/7] 3. (S) The following algorithm outlines the encoding process: [3/5]

116

Appendix B: Exam solutions I

1. Initialise the Huffman tree T containing the only node DAG. 2. while (more characters remain) do s:= next_symbol_in_text(); if (s has been seen before) then output h(s) else output h(DAG) followed by g(s) T:= update_tree(T); en d The function update tree 1. If s is not in S, then add s to S; weight_s:=1; else weight_s:=weight_s + 1; 2. Recompute the Huffman tree for the new set of weights and/or symbols. 4. Example: (Please note that a different but correct process is possible) [5]
For example, let g(s) be an ASCII code for symbol s and h(s) be the Huffman code for symbol s (this is available from the Huffman tree on each step). (0) Initialise S={DAG(0)} Tree (with a single node): DAG(0) (1) Read-input: Output : : S={C(1), DAG(0)} Tree: 1 / C(1) DAG(0) (2) Read-input: Output : : A h(DAG) ASCII("A") i.e. 1 ASCII("A") \

[2/5]

C h(DAG) ASCII("C") i.e. 1 ASCII("C") [or 0 ASCII("C")]

S={C(1), A(1), DAG(0)} Tree: 2 / \ 1 C(1) / \ A(1) DAG(0) (3) Read-input: A Output : 00 S={A(2), C(1), DAG(0)}

117

Data Compression

Tree: / A(2)

3 \ 1 / \ C(1) DAG(0)

(4) read-input: Output :

A 0

A={A(3), C(1), DAG(0)} Tree: 4 / \ A(3) 1 / \ C(1) DAG(0) (5) read-input: Output : B h(DAG) ASCII("B")

A={A(3), C(1), B(1), DAG(0)} Tree: 5 / \ A(3) 2 / \ 1 C(1) / \ B(1) DAG(0) (6) read-input: B Output : 100 A={A(3), B(2), C(1), DAG(0)} Tree: 6 / \ 3 A(3) / \ B(2) 1 / \ C(1) DAG(0)

118

Appendix B: Exam solutions I

Question

A.2 [1/4 each]

1. (S) A can be separated into 4 bitplanes as follows: 0110 1110 1111 0000 1011 0100 1001 0001 0000 1000 0000 0011 2. (UB) [4] [3] [1]

[2]

A Huffman code is optimal only if the average (expected) length of the code is equal to the entropy of the source. Consider a code with n codewords. Let the lengths of codewords of the code are l1 , l2 , .., ln . The average length of the codewords is: l=
n X

pi i=1 where pi is the probability of the codeword i. The entropy of the code is: H =
n X i=1

1 pi log2 . pi

[2/5] P In comparison of the two formula, we find there are n items in both s, so l = H iff li = log2 p = log2 pi , for i = 1, 2, . . . , n.
1
i

Therefore, pi = 2li , li is a non-negative integer. As we can see, the probabilities have to be a negative power of 2. An example (such as the case when n = 4). [2/5] [1/5]

119

Data Compression

3. (S) Linit: A B C D E F 34 25 13 12 9 7 Lcomb: empty B C D 34 25 13 12

Linit: A

Lcomb: EF 16 Linit: A B 34 25 Lcomb: CD EF 25 16

Linit: A 34 Lcomb: B[EF] CD 41 25 Linit: empty Lcomb: A[CD] B[EF] 59 41 Linit: empty Lcomb: [A[CD]] [B[EF]] 10 0 [2/8] [A[CD]][B[EF]] / \ A[CD]] B[EF] / \ / \ A CD B EF / \ / \ C D E F [2/8] Assigning 0 to the left branch and 1 to the right. The canonical and minimum variance Huffman code is A 00 C 010 D 011 B 10 E 110 F 111 [2/8] [2/8]

4. (UB) (a) This is a model where a fixed number of symbols (usually one symbol) of the alphabet are assigned to a codeword and the code for the alphabet contains codewords of different length. In other words, the number of bits of the codeword for each group of fixed number 120

Appendix B: Exam solutions I

of symbols is a variable, i.e. the numbers of bits used for coding these units of fixed number of symbols are different. [2/8] For example, Huffman coding belongs to this class. It assigns a longer codeword to the symbol that occurs less frequently in the source file, and a shorter codeword to the one that occurs more frequently. [2/8] (b) Gray-scale image is a type of image with 8 bit colour depth. The images are captured by scanners using multiple intensity levels to record shadings between black and white. Each pixel is represent ed by 8 bits binary number which hold one colour value to provide 256 different shades of grey. [2/8] Medical images and black and white photos are examples for gray scale images. An image consists of pixels and each pixel of the image is represented by a 8-bit shade value. For example, 00001000 can be used to represent a pixel with the shade of the value. [2/8]

121

Data Compression

Question

A.3

1. (U) The colour depth is represented by the number of bits used to hold a colour value in a RGB colour model. [2/4] The value of the colour depth in this case: 8. [2/4] 2. (S) Compressing a file can be viewed as mapping it to a different (hopefully a shorter) file. Compressing a file of n bytes (in size) by at least 1 byte means mapping it to a file of n 1 or fewer bytes. The re are 256n files of n bytes and 256 n1 of n 1 bytes in total. This means the mapping proportion is only 256n1 /256n = 1/256 which is less than 1%. [4] 3. (U) If we follow the rules below during the encoding process, only one Huffman code will be derived and the code is called canonical. Also, the codewords of the code will have the minimum variation in length. i) A newly created element is always placed in the highest position possible in the sorted list. ii) When combining two items, the one higher up in the list is assigned a 0 and the lower down is a 1. Huffman coding that follows these rules is called Canonical and Minimum variance Huffman coding. [2/4]

It would be possible to derive two different Huffman codes for the same probability distribution of an alphabet, because there can be items with equal probabilities at some stage on the sorted probability/f requency list, and this leads to two (or more) possible positions in which the symbols involved are placed in the tree, [1/4] and the roles of 0 and 1 are interchangeable depending on our rules. [1/4] 4. (U) The Shannon- Fano encoding and static Huffman coding algorithms both begin by creating one node for each character, but they build the tree differently. The Huffman method starts from the bottom (leaves) and builds the tree in a bottom-up approach, while The ShannonFano method starts from the top (root) and builds the tree in a topdown approach. [2/5] Example: {A,B,C, D} with probability distribution 0.5, 0.3, 0.1, 0.1. [3/5] Huffman encoding: A B C D 0.5 0.3 0.1 0.1 B CD 0.5 0.3 0.2 A

122

Appendix B: Exam solutions I

B[CD] A 0.5 0.5 [B[CD]]A 1 Huffman tree: [B[CD]]A / \ B[CD] A / \ B CD /\ C D Shannon- Fano encoding A [ 0.5 [ A [ 0.5 [ A [ 0.5 [ B C D ] 0.3 0.1 0.1 ] B [C D ]] 0.3 [0.1 0.1]] B 0.3 [ C [ 0.1 D]] 0.1]

A[B[CD]] Shannon-Fano tree: A[B[CD]] / \ A B[CD] / \ B CD / C

\ D

123

Data Compression

5. (U)

[(1/2+1/2+1)/8

each, for figure, prefix code and explanation] r 0/ \1 0/ C 0/ \1 0/\1 D A B (III) r 0/ \1 0/ D 0/ 0/\1 B C,A (IV)

r 0/ \1 0/ \1 0/\1 0/\1 A BC D (I)

r 0/ \1 C D \1 0/\1 A B (II)

(I) is a prefix code, because all the symbols are at the leaves and no symbols share any node. (II) is not a prefix code, because D is at an internal node. This means that the code for D is a prefix of A and B. (III) is a prefix code, because all the symbols are at the leaves and no symbols share any node. (IV) is not a prefix code, because A and C are represented by the same code (a prefix code should be uniquely decodable) which cannot be uniquely decodable. A and C share the same leaf although all symbols are leaves.

124

Appendix B: Exam solutions I

Question 1. (B)

A.4

(a) rate-distortion problem This is one of the two types of data compression problems (proposed by Davisson and Gray 1976). The compression problems of this type are to meet the requirement of achieving certain pre-specified fidelity with the best bit-rate possible, i.e. as few bits per second as possible. [2/8] (b) entropy The theoretical minimum average number of bits that are required to transmit a particular source stream is known as the entropy of the source and can be computed using Shannons formula (in 1940s): 1 n X H = pi log2 pi i=1 [2/8] (c) prefix codes A prefix code is a code in which no codeword is a prefix to another codeword. This is when no sequence of bits which codes for one symbol is a prefix of the codeword for another symbol. [2/8] (d) motion prediction This is a class of basic compression techniques which are mainly for lossy image compression. The idea is to replace objects (e.g. a 8 8 block of pixels) in frames with references to the same object (at a slightly different position ) in the previous frame. [2/8] 2. (S) Since l1 = 1, l2 = 2, and l3 = l4 = 3. We have
4 X

= 1/2 + 1/22 + 1/23 + 1/23 = 1/2 + 1/4 + 1/8 + 1/8 [2/4]

=1
i=1

As we see the Kraft-MacMi llians inequation is satisfied, so Yes, it is possible to find a uniquely decodable code for these lengths. [2/4] 3. (S) The output would be: n3 TTUr9 n1 Kr2 n2 RRr14 3r2 r3 Pn2 EE [2/5] Explanation of the control characters used: n3 TTU: 3 non-repeating characters TTU. r9 : 9 repeating blanks. n1 K: 1 non-repeating characters K. r2 : 2 repeating blanks. n2 RR: 2 non-repeating characters RR. 125 [1/5]

Data Compression

r14 3: 14 repeating 3s. r2 : 2 blanks. r3 P: 3 repeating Ps. n2 EE: 2 non-repeating characters EE. Possible data structu re(s) and algorithm(s): Array, List, Tree, and a simple correct counting algorithm. [2/5]

4. (BS) These are also called photographic images for 2-dimensional pictures. A bitmapped image is an array of pixel values. The data are the values of pixels. Many bitmapped images are created from digital devices such as scanners or digital cameras, or from programs such Painter which allows visual artists to paint images. [2/8] The image in vector graphics is stored as a mathemati cal description of a collection of graphic components such as lines, curves and shapes etc.. Vector graphics are normally more compact, scalable, resolution independent and easy to edit. They are very good for 3-D models on computer, usually built up using shapes or geometric objects that can easily be described mathemati cally. [2/8] The [requirements] to the computer system: a bitmapped image must record the value of every pixel, but vector description takes much less space in an economy way. [2/8] The [size] of bitmapped images depends on the size of image and the resolution but is independent from the complex of the image; while the size of vector graphics depends on the number of objects consist of the image, but is independent of any resolution. [2/8]

126

Appendix B: Exam solutions I

Question

A.5

1. (U) Because we can always use the fixed-length to decode a sequence of zeros and ones. [2/4] For example, 01000100001001000100 can be easily divided into codewords if we know that each codeword is 4 bits long: 0100,0100,0010,0100,0100. [2/4] 2. The main idea of predictive encoding: is to predict the values of a pixel based on previous pixels and storing the difference between the predicted and actual values in a residual image. [2/6] (a) Derive the residual matrix: 1 0 0 0 5 -4 0 0 5 0 0 0 7 2 -5 1 [2/6]

(b) Use a prefix coding approach, such as Huffman coding (or any suitable compression technique): [2/6] Frequency table: ----------------5 1 -4 1 0 8 1 2 2 1 5 2 7 1

127

Data Compression

3. (U) (a) Encoding: [6/15]

word x (Output-token) Dictionary ------------------------------------------------------------------1 A /* standard entries */ 2 B ... 256 / 0 /* initialisation */ .................................................................. 1 B B .................................................................. 2 B 2 257 BB B .................................................................. 3 G 2 258 BG G .................................................................. 4 B 7 259 GB B .................................................................. 5 G BG .................................................................. 6 H 258 260 BGH H .................................................................. 7 8 /* end of the algorithm */ ------------------------------------------------------------------

So the decoded string is 2 2 7 258 8.

128

Appendix B: Exam solutions I

(b) Decoding

[7/15]

word x (Output) Dictionary --------------------------------------------------------------------1 A /* standard entries */ 2 B . . 256 / .................................................................. 0 2 B B .................................................................. 1 2 B 257 BB B .................................................................. 2 7 G 258 BG G .................................................................. 3 258 BG 259 GB BG .................................................................. 4 8 H 260 BGH H /* end of the algorithm */ ------------------------------------------------------------------

So the decoded string is BBGBGH. (c) Conclusion: [2/15] As we can see, the dictionary produced during encoding and the one during decoding are identical. Each dictionary is built independently without the need for transmission.

129

Data Compression

130

Appendix C: Sample examination paper II

Ap pendix C

Sample examination
Data Com pression
Duration: 2 hours and 15 minutes

pa per II

You must answer FOUR questions only. The full mark for each question is 25. The marks for each subquestion are displayed in brackets (e.g. [3]). Electronic calculators may be used. The make and model should be specified on the script and the calculator must not be programmed prior to the examination. Question C.1

1. Explain the meaning of the following terms: [8] (a) rate-distortion problem (b) entropy (c) prefix codes (d) motion prediction. 2. Explain the concept of the following and provide one example for each case. [8] (a) fixed-to-variable model (b) gray-scale image 3. It has been suggested that unique decodability is not a problem for any fixed length codeword set. Explain why this is so with the aid of an example. [4] 4. Provide reasons, with the aid of an example, to support the following statement: [5] The Huffman code in general is not optimal unless all probabilities are negative powers of 2.

131

Data Compression

Question

C.2

1. Comment on the truth of the following statement in describing the absolute limits on lossless compression. [4] No algorithm can compress even 1% of all files, even by one byte. 2. Explain briefly what is meant by Canonical and Minimum-varian ce Huffman coding, and why it is possible to derive two different Huffman codes for the same probability distribution of an alphabet. [4] 3. Describe briefly, with the aid of an example, how Shannon- Fano encoding differs from static Huffman coding. [5] Hint: You may give an example by deriving a code for {A,B,C, D} with probability distribution 0.5, 0.3, 0.1, 0.1 using the two algorithms, and show the difference. 4. A binary tree (0-1 tree) can be used to represent a code containing a few codewords of variable length. Consider each of the four codes for characters A,B,C,D below and draw the binary trees for each code. (a) {000, 001, 110, 111} (b) {110,111,0, 1} (c) {0000,0001,1,00 1} (d) {0001,0000,0001, 1} For each tree drawn, comment on whether the code being represented by the binary tree is a prefix code and give your reasons. [12]

132

Appendix C: Sample examination paper II

Question

C.3

1. Given an alphabet of four symbols {A,B,C, D}, discuss the possibility of a uniquely decodable code in which the codeword for A has length 1, that for B has length 2 and for both C and D have length 3. Give your reasons. [4] 2. What would be the output if the HDC algorithm is applied to the sequence below? Explain the meaning of each control symbol that you use. [5] TTUKRR33333333333333PPPEE 3. One way to improve the efficiency of Huffman coding is to maintain two sorted lists during the encoding process. Using this approach, derive a canonical minimum variance Huffman code for the alphabet {A,B,C,D,E, F} with the probabilities (in %) 34,25,13,12,9,7 respectively. [8] 4. Explain the concept of bitmapped images and vector graphics. What are the differences between vector and bitmapped graphics in terms of requirements to the computer storage capacity, and the size of the image files. [8]

133

Data Compression

Question

C.4

1. Determine whether the following codes for the alphabet {A, B, C, D} are uniquely decodable. Give your reasons for each case. [8] (a) {1, 10, 101, 0101} (b) {000, 001, 010, 111} (c) {0, 001, 10, 011} (d) {000, 010, 011, 1} 2. Consider a source with a binary alphabet {B, W}. Under what condition on the source does the static Huffman coding algorithm achieve the best performance? Under what condition on the source does the algorithm perform the worst? Give your reasons. [7] 3. Illustrate how adaptive Huffman encoding works. [6] Hint: You may demonst rate step by step the progress of running an Adaptive Huffman encoding algorithm in terms of the input, output , alphabet and the tree structure. Suppose the sequence of symbols to be encoded from the initial state is CAAABB. 4. Explain what is used to represent the so-called colour depth in a common RGB colour model. What is the value of the colour depth in a representation where 8 bits is assigned to every pixel? [4]

134

Appendix C: Sample examination paper II

Question

C.5

1. Huffman coding can perform badly when it is applied to fax images. For example, the canonical minimum-varian ce Huffman code is about 37% worse than its optimal when the probabilities of the white pixels P r(W hite) = 0.2 and P r(Black) = 0.8. Demonst rate step by step how this situation can be improved by blocking two symbols at a time from the source. [12] Hint: log10 2 0.3; log10 0.8 0.1; log10 0.2 0.7. 2. Derive the reflected Gray code for decimal number 5. 3. A binary sequence of length 5 (symbols) was encoded on the binary alphabet {A,B } using Arithmetic coding. Suppose the probability P r(A) = 0.2. If the encoded output is 0.24, derive the decoded output step by step. [10] The decoding algorithm is given below. 1. (Initialise) L <- 0 and d <- 1 2. (Probabilities) For the next symbol, Let Pr[s1] be p1, the probability for s1 and Pr[s2] be p2, the probability for s2 3. (Update) If x is a member of [L,L+d*p1) then output s1, leave L unchanged, and set d<-d*p1 else if x is a member of [L+d*p1, L+d) then output s2, set L<- L+d*p1 and d<-d*p2 4. (Done?) If the_number_of_decoded_symbols < the_required_number_of_symbols then go to step 2. [3]

135

Data Compression

Question

C.6

1. Consider part of a grayscale image with 16 shades of gray that is represented by the array A below: 001 100 100 001 110 100 110 011 1 0 0 0 0 100 0 110 0 100 0 100 0 0 1 1 Demonst rate how the image can be represented by several bitplanes (bi-level images)

[4]

2. Describe the main idea of predictive encoding. Suppose the matrix below represents the pixel values (in decimal) of part of a grayscale image. Let the prediction be that each pixel is the same as the one to its left. Illustrate step by step how predictive encoding may be applied to it. [6] 1 5 5 7 1 1 5 9 1 1 5 4 1 1 5 5

3. Demonst rate step by step how the Basic LZW encoding and decoding algorithms maintain the same version of a dictionary without ever transmitting it in a separate file, using a small string AABABC as an example. [15] The LZW encoding and decoding algorithms are given below. (a) Encoding w o rd = ; while not end_of_file x=read_next_character ; if word+x is in the dictionary word=word+x else output the dictionary index for word; add word+x to the dictionary; word= x; output the dictionary number for word; (b) Decoding read a token x from compressed file; output element word=eleme nt; while not end_of_compressed_file do read x; look up dictionary for element at x; if there is no entry yet for index x then element=word+first_char_of_word; output element; add word+first_char_of_element to the dictionary; word=eleme nt;

136

Appendix D: Exam solutions II

Ap pendix D

Exam solutions II
Please be aware that these solutions are given in brief note form for the purpose of this section. Candidates for the exam are usually expected to provide coherent answers in full form to gain full marks. Also, there are other correct solutions which are completely acceptable but which are not included here. Duration: 2 hours and 15 minutes You must answer FOUR questions. The full mark for each question is 25. The marks for each subquestion are displayed in brackets (e.g. [3]). Electronic calculators may be used. The make and model should be specified on the script and the calculator must not be programmed prior to the examination. Question 1. (B) (a) rate-distortion problem This is one of the two types of data compression problems (proposed by Davisson and Gray 1976). The compression problems of this type are to meet the requirement of achieving certain pre-specified fidelity with the best bit-rate possible, i.e. as few bits per second as possible. [2/8] (b) entropy The theoretical minimum average number of bits that are required to transmit a particular source stream is known as the entropy of the source and can be computed using Shannons formula (in 1940s): 1 n X H = pi log2 pi i=1 [2/8] (c) prefix codes A prefix code is a code in which no codeword is a prefix to another codeword. This is when no sequence of bits which codes for one symbol is a prefix of the codeword for another symbol. [2/8] (d) motion prediction This is a class of basic compression techniques which are mainly for lossy image compression. The idea is to replace objects (e.g. a 8 8 block of pixels) in frames with references to the same object (at a slightly different position ) in the previous frame. [2/8] 2. (UB) (a) This is a model where a fixed number of symbols (usually one symbol) of the alphabet are assigned to a codeword and the code for the alphabet contains codewords of different length. In other words, 137 A.1

Data Compression

the number of bits of the codeword for each group of fixed number of symbols is a variable, i.e. the numbers of bits used for coding these units of fixed number of symbols are different. [2/8] For example, Huffman coding belongs to this class. It assigns a longer codeword to the symbol that occurs less frequently in the source file, and a shorter codeword to the one that occurs more frequently. [2/8] (b) Gray-scale image is a type of image with a 8 bit colour depth. The images are captured by scanners using multiple intensity levels to record shadings between black and white. Each pixel is represent ed by 8 bits binary number which hold one colour value to provide 256 different shades of grey. [2/8] Medical images and black and white photos are examples for gray scale images. An image consists of pixels and each pixel of the image is represented by a 8-bit shade value. For example, 00001000 can be used to represent a pixel with the shade of the value. [2/8] 3. (U) Because we can always use the fixed-length to decode a sequence of zeros and ones. [2/4] For example, 01000100001001000100 can be easily divided into codewords if we know that each codeword is 4 bits long: 0100,0100,0010,0100,0100. 4. (UB) A Huffman code is optimal only if the average (expected) length of the code is equal to the entropy of the source. Consider a code with n codewords. Let the lengths of codewords of the code are l1 , l2 , .., ln . The average length of the codewords is l=
n X

[2/4]

pi
i= 1

where pi is the probability of the codeword i. The entropy of the code is

n X 1 H = pi log2 pi i= 1

[2/5]

In comparison of the two formula, we found there are n items in both P s), So l = H iff li = log2 p 1 = log2 pi , for i = 1, 2, . . . , n.
i

So pi = 2li , li is a non-negative integer. As we can see, the probabilities have to be a negative power of 2. An example (such as the case when n = 4). [2/5] [1/5]

138

Appendix D: Exam solutions II

Question

A.2

1. (S) Compressing a file can be viewed as mapping it to a different (hopefully a shorter) file. Compressing a file of n bytes (in size) by at least 1 byte means mapping it to a file of n 1 or fewer bytes. The re are 256n files of n bytes and 256 n1 of n 1 bytes in total. This means the mapping proportion is only 256n1 /256n = 1/256 which is less than 1%. [4] 2. (U) If we follow the rules below during the encoding process, only one Huffman code will be derived and the code is called canonical. Also, the codewords of the code will have the minimum variation in length. i) A newly created element is always placed in the highest position possible in the sorted list. ii) When combining two items, the one higher up in the list is assigned a 0 and the lower down is a 1. Huffman coding that follows these rules is called Canonical and Minimum variance Huffman coding. [2/4] It would be possible to derive two different Huffman codes for the same probability distribution of an alphabet, because there can be items with equal probabilities at some stage on the sorted probability/f requency list, and this leads to two (or more) possible positions in which the symbols involved are placed in the tree, [1/4] and the roles of 0 and 1 are interchangeable depending on our rules. [1/4] 3. (U) Shannon- Fano encoding and static Huffman coding algorithms both begin by creating one node for each character, but they build the tree differently. Huffman method starts from the bottom (leaves) and builds the tree in a bottom-up approach, while Shannon- Fano method starts from the top (root) and builds the tree in a top-down approach. [2/5] Example: {A,B,C, D} with probability distribution 0.5, 0.3, 0.1, 0.1. [3/5] Huffman encoding: A B C D 0.5 0.3 0.1 0.1 A B CD 0.5 0.3 0.2 B[CD] A 0.5 0.5 [B[CD]]A Huffman tree: [B[CD]] A / \ B[CD] A / \ B CD /\ C D 139

Data Compression

Shannon- Fano encoding A [ 0.5 [ A [ 0.5 [ A [ 0.5 [ B C D ] 0.3 0.1 0.1 ] B [C D ]] 0.3 [0.1 0.1]] B 0.3 [ C [ 0.1 D]] 0.1]

A[B[CD]] Shannon-Fano tree: A[B[CD]] / \ A B[CD] / \ B CD / C 4. (U) r 0/ \1 0/ \1 0/\1 0/\1 A BC D (I)

\ D

[(1+1+1)/12 each, for figure, prefix code and explanation] r 0/ \1 C D \1 0/\1 A B (II) r 0/ \1 0/ C 0/ \1 0/\1 D A B (III) r 0/ \1 0/ D 0/ 0/\1 B C,A (IV)

(I) a prefix code, because all the symbols are at the leaves and no symbols share any node. (II) not a prefix code, because D is at an internal node. This means that the code for D is a prefix of A and B. (III) a prefix code, because all the symbols are at the leaves and no symbols share any node. (IV) not a prefix code, because A and C are represented by the same code (a prefix code should be uniquely decodable) which cannot be uniquely decodable. A and C share the same leaf although all symbols are leaves.

140

Appendix D: Exam solutions II

Question 1. (S)

A.3

Since l1 = 1, l2 = 2, and l3 = l4 = 3. We have

4 X

= 1/2 + 1/22 + 1/23 + 1/23 = 1/2 + 1/4 + 1/8 + 1/8 [2/4]

=1
i=1

As we see the Kraft-MacMi llians inequation is satisfied, so Yes, it is possible to find a uniquely decodable code for these lengths. [2/4] 2. (S) The output would be: n3 TTUr9 n1 Kr2 n2 RRr14 3r2 r3 Pn2 EE [2/5] Explanation of the control characters used: n3 TTU: 3 non-repeating characters TTU. r9 : 9 repeating blanks. n1 K: 1 non-repeating characters K. r2 : 2 repeating blanks. n2 RR: 2 non-repeating characters RR. r14 3: 14 repeating 3s. r2 : 2 blanks. r3 P: 3 repeating Ps. n2 EE: 2 non-repeating characters EE. 3. (S) Linit: A B C D E F 34 25 13 12 9 7 Lcomb: empty Linit: A B C D 34 25 13 12 Lcomb: EF 16 Linit: A B 34 25 Lcomb: CD EF 25 16 Linit: A 34 Lcomb: B[EF] CD 41 25 Linit: empty Lcomb: A[CD] B[EF] 141 [3/5]

Data Compression

Linit: empty Lcomb: [A[CD]] [B[EF]] 10 0 [2/8] [A[CD]][B[EF]] / \ A[CD]] B[EF] / \ / \ A CD B EF / \ / \ C D E F [2/8] Assigning 0 to the left branch and 1 to the right. The canonical and minimum variance Huffman code is A 00 C 01 D0 01 B1 10 E 11 F 0 11 1 4. (BS) These are also called photographic images for 2-dimensional pictu re. A bitmapped image is an array of pixel values. The data are the values of pixels. Many bitmapped images are created from digital devices such as scanners or digital cameras, or from programs such Painter which allows visual artists to paint images. [2/8] The image in vector graphics is store as a mathemati cal description of a collection of graphic components such as lines, curves and shapes etc.. Vector graphics are normally more compact, scalable, resolution independent and easy to edit. They are very good for 3-D models on computer, usually built up using shapes or geometric objects that can easily be described mathemati cally. [2/8] The [requirements] to the computer system: a bitmapped image must record the value of every pixel, but vector description takes much less space in an economy way. [2/8] The [size] of bitmapped images depends on the size of image and the resolution but is independent from the complex of the image; while the size of vector graphics depends on the number of objects consist of the image, but is independent of any resolution. [2/8] [2/8] [2/8]

142

Appendix D: Exam solutions II

Question 1. (S)

A.4

(a) No. Because, for example, 01011010101 can be decoded either as DCD or as DBBC (or DBAD). [1+1/8 ] (b) Yes, because all the codewords have the same length. [1+1/8 ] (c) No, because for example, 00100010 can be decoded as BAA AC or AACBA . [1+1/8 ] (d) Yes, because this is a prefix code. 2. (U) Let the probability distribution of the source be {pB , pW }. [1+1/8 ] [1/7]

When pB = pW = 1/2, the static Huffman algorithms work the best. [1/7] This is because that the average length of the Huffman code is equal to the entropy of source. The Huffman code is optimal. The average length of the codewords is 1 and the entropy of the source is 1/2( log2 1/2)1/2( log2 1/2)+ = 1. [2/7] When pB = 0, pW = 1 (or pB = 1, pW = 0), the static Huffman algorithms work the worst. [1/7]

The average length of the codewords is 1 and the entropy of the source is 1( log2 1) = 0. [2/7] 3. (S) The following algorithm outlines the encoding process: 1. Initialise the Huffman tree T containing the only node DAG. 2. while (more characters remain) do s:= next_symbol_in_text(); if (s has been seen before) then output h(s) else output h(DAG) followed by g(s) T:= update_tree(T); en d The function update tree 1. If s is not in S, then add s to S; weight_s:=1; else weight_s:=weight_s + 1; 2. Recompute the Huffman tree for the new set of weights and/or symbols. [1/6]

143

Data Compression

Example: (Please note that a different but correct process is possible) [5/6] For example, let g(s) be an ASCII code for symbol s and h(s) be the huffman code for symbol s (this is available from the Huffman tree on each step). (0) Initialise S={DAG(0)} Tree (with a single node): DAG(0) (1) Read-input: Output : : S={C(1), DAG(0)} Tree: 1 / C(1) DAG(0) (2) Read-input: Output : : A h(DAG) ASCII("A") i.e. 1 ASCII("A") \

C h(DAG) ASCII("C") i.e. 1 ASCII("C") [or 0 ASCII("C")]

S={A(1), C(1), DAG(0)} Tree: 2 / \ A(1) 1 / \ C(1) DAG(0) (3) Read-input: A Output : 00 S={A(2), C(1), DAG(0)} Tree: 3 / \ A(2) 1 / \ C(1) DAG(0) (4) read-input: Output : A 0

A={A(3), C(1), DAG(0)}

144

Appendix D: Exam solutions II

Tree: / A(3)

4 \ 1 / \ C(1) DAG(0)

(5) read-input: Output :

B h(DAG) ASCII("B")

A={A(3), C(1), B(1), DAG(0)} Tree: 5 / \ A(3) 2 / \ 1 C(1) / \ B(1) DAG(0) (6) read-input: B Output : 100 A={A(3), B(2), C(1), DAG(0)} Tree: 6 / \ 3 A(3) / \ B(2) 1 / \ C(1) DAG(0) 4. (U) The colour depth is represented by the number of bits used to hold a colour value in a RGB colour model. [2/4] The value of the colour depth in this case: 8. [2/4]

145

Data Compression

Question

A.5

1. (U) This situation can be improved by the approaches below: (a) By grouping 2 symbols at a time, the binary alphabet is extended to {AA, AB, BA, BB} which consists of four elements. [2/12] (b) (Sub-total for this part The probabilities for the combined symbols are: A 0.6 A 0.1 A 4 B 6 0.1 A 6 B 0.0 B 4 AA AB BA BB 0.64 0.16 0.16 0.04 AA [BA BB] AB 0.64 0.2 0.16 AA [BA BB] AB 0.64 0.2 0.16 AA [[BA BB] AB] 0.64 0.36 [AA [[BA BB] AB]] 1 [2/12] [AA [[BA BB] AB]] / \ AA [[BA BB] AB] / \ [BA BB] AB / \ BA BB [2/12] [8/12]) [2/12]

146

Appendix D: Exam solutions II

So the Huffman code: AA BB BA BB 0 11 10 0 10 1

[2/12] 1 0.64 + 2 0.16 + 3 0.16 + 3 0.04 = 1.56bits

(c) Now the average length is, for two symbols,

This is 0.78 bits/sym bol which is closer to the entropy 0.72 bit/sym bol compared with before: average length was 1 bit/sym bol. [2/12] 2. (UB) The binary code of 5 is 101. Shift 101 one bit to the right: 010 The reflected Gray code for 5 is: 101 XOR 010 = 111 3. (S) Let s1=A, s2=B, p1 = 0.2, p2 = 0.8 and x = 0.12. The following table shows the values of main variables in each round. [8/10]
L 0 1 2 0.2 3 0.232 4 0.232 5 0.23712 0.0189696 0.0256 0.00512 0.02048 [0.232, 0.23712) [0.23712, 0.2576) B 0.128 0.0256 [0.232, 0.2576) A 0.16 0.032 0.128 [0.2, 0.232) [0.232, 0.36) B 0 0.2 d 1 0.8 0.16 dp1 0.2 0.8 [0.2, 0.36) A dp2 [L, L+dp1) [0, 0.2) [L+dp1, L+d ) [0.2, 1) O utput B

[1/3] [1/3] [1/3]

So the decoded string is BABAB.

[2/10]

147

Data Compression

Question

A.6 [1/4 each]

1. A can be separated into 4 bitplanes as following: 0110 1110 1111 0000 1011 0100 1001 0001 0000 1000 0000 0011 [4] [3] [1]

[2]

2. Main idea of predictive encoding: is to predict the values of a pixel based on previous pixels and storing the difference between the predicted and actual values in a residual image. [2/6] (a) Derive the residual matrix: 1 5 5 7 0 -4 0 2 0 0 0 -5 0 0 0 1 [2/6]

(b) Use a prefix coding approach, such as Huffman coding (or any suitable compression technique): [2/6] Frequency table: ----------------5 1 -4 1 0 8 1 2 2 1 5 2 7 1

148

Appendix D: Exam solutions II

3. (U) (a) Encoding:

word x (Output-token) Dictionary --------------------------------------------------------1 A /* standard entries */ 2 B ... 256 / 0 /* initialisation */ ......................................................... 1 A A ......................................................... 2 A 1 257 AA A ......................................................... 3 B 1 258 AB B ......................................................... 4 A 2 259 BA A ......................................................... 5 B AB ......................................................... 6 C 258 260 ABC C ......................................................... 7 3 /* end of the algorithm */ ---------------------------------------------------------

[6/15]

So the decoded string is 1 1 2 258 3.

149

Data Compression

(b) Decoding
word x (Output) Dictionary -----------------------------------------------------1 A /* standard entries */ 2 B . . 256 / ...................................................... 0 1 A A ...................................................... 1 1 A 257 AA A ...................................................... 2 2 B 258 AB B ...................................................... 3 258 AB 259 BA AB ...................................................... 4 3 C 260 ABC C /* end of the algorithm */ ------------------------------------------------------

[7/15]

So the decoded string is AABABC. (c) Conclusion: [2/15] As we can see, the dictionary produced during encoding and the one during decoding are identical. Each dictionary is built independently without the need for transmission.

150

Appendix E: Support reading list

Ap pendix E

Sup port reading list

This is a collection of textbooks in addition to those recommended at various points in the subject guide. It is for students further interests on the subject and includes some historical backgrounds and useful web sites.

Text books
Drozdek, Adam Elements of Data Compression. (Brooks/Cole, USA, 2002) [ISBN 0-534-38448-X]. Gibson, Jerry D., Berger, Toby, Lookabaugh, Tom, Lindbergh, Dave, Baker, Richard L. Digital Compression for Multimedia : Principles and Standa rds. (San Francisco: Morgan Kaufmann Publishers, 1998) [ISBN 1-55860-369-7]. Nelson, Mark and Gailly, Jean-Loup The Data Compression Book. 2nd edition (New York: M&T Books, 1996) [ISBN 1-55851-434-1]. Ross, Sheldon, A First Course in Probability 3rd edition (New York: MacMillan, 1988) [ISBN 0-02-403850-4]. Sayood, Khalid Introduction to Data Compression (San Diego: Morgan Kaufmann, 1996) [ISBN 1-55860-346-8].

Web addresses Data Compression

http://www.ics.uci.edu/~dan/pubs/DataCompression.ht ml

151

Data Compression

Index
2-dimensional, 69 adaptive, 24 Huffman, 42 ADPCM, 97 advantages, 83 AIFF, 98 alpha bet, 27, 43 source, 66 amplitude, 95, 97 Animated, 74 appr oximate statistical data, 17 arithmetic codes, 45 coding, 43 art, 86 ASCII, 13, 39 attribute, 28 AU, 98 average, 28, 29 B-pictures, 91 balanced, 17 basis, 85 basis vector, 84 behaviour, 71 binary alpha bet, 44 binary tree, 14, 15 bitmap images, 69 BMP, 74 buffer History buffer, 60 Lookahead buffer, 60 canonical, 17 canonical minimum-variance, 17 cardinali ty, 27 changed, 43 chrominance, 72, 73 code, 13 coder, 23, 43, 51 codeword, 13, 26 codewords, 19 coefficients, 85, 93 colour, 71 database, 72 depth, 72 communication 152 channel, 83 compression, vii, 1 lossy, 82 music, 97 ratio, 97 voice, 97 compression ratio, 17 compressor, 39 compressors, 51 context, 78 converge, 28 correlated pixels, 84 curves, 70 dangling, 26 DAT, 96 data, 2 DCT, 86 decimal digits, 43 number, 44, 46, 47 value, 49 decimal system, 47 decodable, 13 decoder, 23, 43, 45, 56 decompressor, 39, 40 decompressors, 51 decorrelated, 84 descending, 14 detect, 95 device resolution, 70 dictionar y, 51, 52, 59 dictionary-based algorithms, 51 approaches, 51 digital, 96 audio, 98 cameras, 69 devices, 69 images, 69 distribution, 26 disturbance, 95 division, 18, 19, 45 dpi, 69 drawing, 71 EBCDIC, 13

Index

efficiency, 7 encoder, 23, 43, 45 energy, 85 entire input, 43 entropy, 28, 81 entropy of the source, 28 entropy-based, 51 EPS, 74 estimation, 81 event, 27, 28 events, 97 expect, 28 extending alpha bet, 41 fixed, 25 probabili ty distribution, 43 fixed-length, 13 fixed-tovariable, 51 frame, 89 difference, 91 rate, 89 frames, 91 frequency, 39 spectrum, 96 frequency spectrum diagram, 96 frequency table, 14 FT, 86 GIF, 74, 79 goal, 44 grayscale, 78 grouping symbols, 41 HDC, 9 HDTV, 89 hierar chical coding, 84 HT, 86 Huffman, 13, 23, 39 Huffman code, 16 Huffman tree, 14, 16 I-picture, 91 image, 2 bi-level, 73 cartoonlike, 74 colour, 73 compression, 77 continuous-tone, 74 discrete-tone, 74 files, 69 graphical, 74 gray-scale, 73 progressive, 83

reconstructed, 82 synthetic, 74 images bi-level, 79 impleme nt, 83 impleme ntation problem, 48 indices, 51 information source, 27 information theory, 27 instructions, 97 interval, 43, 44, 46, 48, 96 intervals, 91 intra, 91 ITU-T, 97 ITU-T H.261, 90 JPEG, 74, 81, 84 Still, 86 KLT, 86 Kraft-McMillan, 27 large alphabets, 48 LC, 72 leaf, 14, 16 lifetime distribution, 44 lightness, 72 lines, 70 lossless, 2, 29, 77 lossy, 2, 77, 86, 97 compression, 69 luminance, 72 LZ77, 51 LZ78, 51 LZW, 51 match, 60, 62 matrix, 82, 84 measure, 82 difference distortion, 82 medium, 95 memoryless, 44 middle C, 95 MIDI, 99 minimum-variance, 17 mode baseline sequential, 86 lossless, 86 model, 23, 39, 43, 95 MP3, 98 lo-fi MP3, 99 MPEG, 90 multimedia, 2, 97 noises, 97 non-prefix, 26 153

Data Compression

non-symmetric, 24 notes, 95 NTSC, 89 occurrence, 17, 39 offset, 60, 62 optimal, 29 optimum, 29 original, 49 output, 43 P-pictures, 91 painting, 71 PAL, 89 parameters, 89 parity, 13 patterns, 85 photographic image, 69 phrases, 51 pixel context, 78 dimension, 69 pixsel, 2 PNG, 74 portion attack, 95 steady, 95 predict, 98 prediction backward, 93 bidirectional, 93 forward, 92 predictive pictures, 91 predictors, 81 prefix, 26, 60 prefix code, 25, 26 primary colours, 71 priority, 83 probabilities, 39 probabili ty distribution, 43 probabili ty distribution, 30 probabili ty table, 17 probabili ty theory, 27 processing, 85 progressive, 83 PSD, 74 Layered, 75 pyramid coding, 84 quality, 89 quantisation, 96 QuickTime, 99 random variable, 27 range, 96 154

rasterising, 71 rate bit rate, 97 real numbers, 96 RealAudio, 98 RealVideo, 98 reconstructed image, 82 redundanc y, 51, 75 spatial, 90 temporal, 90 rendering, 70 RGB, 72 RGC, 79 root, 14, 16 run-length, 9 sample frequency, 96 samples, 96 sampling, 96, 98 rates, 69 saving percentage, 16 science, 86 SECAM, 89 self-information, 27 sequence, 90 Shannon, 28 Shannon- Fano code, 18 Shannon- Fano coding, 17 shapes, 70 signal, 85, 89 periodic, 95 signals, 95 simple version, 44 sliding window, 59 small alpha bet, 46 Sound, 95 sound, 2, 95 models, 97 source, 17, 43 source alpha bet, 27 speech telephone, 97 wideband, 97 standards, 86 static, 43 still picture, 89 STN, 75 stream, 43 successful, 28 suffix, 26 symmetric, 24 text, 2, 13

Index

compression, 69 stream, 51 TGA, 75 theory Nyquist theory, 96 TIFF, 75 timbre, 95 tokens, 51 trade-off, 86 Transform, 84 transform one-dimensional, 85 two-dimensional, 85 Wavelet, 85 transie nt, 95 traverse, 16 tree, 26 unique interval, 44 uniquely decodable, 29 variable, 25 variable length, 29 variable rate, 96 variablelength, 25 variable-tovariable, 51 vector graphics, 69 vectorisation, 71 video analog, 89 digital, 89 virtual, 97 VOCPACK, 98 WAV, 98 wave continuous wave, 96 waveforms, 97 waves cosine, 95 weight, 39 weighted, 84 WHT, 86 words, 51

155

Data Compression

Notes

156

Notes

157

CIS325Data Compression

Comment form
We welcome any comments you may have on the material s which are sent to you as part of your study pack. Such feedback from student s helps us in our effort to improve the material s produce d for the External Programme. If you have any comments about this guide, either general or specific (includin g corrections, non-availability of essential texts, etc.), please take the time to complete and return this form. Name........................................................................................................................................................................ Address.................................................................................................................................................................... ................................................................................................................................................................................ Em ail ........................................................................................................................................................................ Student number ................................................................................................................. ..................................... For which qualification are you studying? .............................................................................................................. Comments ................................................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ ........................................................................................................................................................ Please continue on additional sheets if necessary. Date: ......................................................................................................................................................................
Please send your comments on this form (or a photocopy of it) to: Publishing Manager, External Programme, University of London, Stewart House, 32 Russell Square, London WC1B 5DN UK.

158

Algorithms On String Trees and Sequences
No ratings yet
Algorithms On String Trees and Sequences
326 pages
Gravis Ultrasound - Installing Your Memory Upgrade
No ratings yet
Gravis Ultrasound - Installing Your Memory Upgrade
2 pages
Large Scale Geometry PDF
No ratings yet
Large Scale Geometry PDF
203 pages
Apt K.R. From Logic Programming To Prolog (PH, 1997) (ISBN 013230368X) (O) (345s) - CSPL
100% (1)
Apt K.R. From Logic Programming To Prolog (PH, 1997) (ISBN 013230368X) (O) (345s) - CSPL
345 pages
Complete Audio Mastering: Practical Techniques
From Everand
Complete Audio Mastering: Practical Techniques
Gebre E. Waddell
5/5 (5)
Clariion Codes
No ratings yet
Clariion Codes
20 pages
Data Compression Techniques
No ratings yet
Data Compression Techniques
21 pages
JPEG Compression Standard
No ratings yet
JPEG Compression Standard
23 pages
Computer Graphics
100% (1)
Computer Graphics
132 pages
Jpeg Image Compression
No ratings yet
Jpeg Image Compression
7 pages
3-0-Fundamental of Compression
No ratings yet
3-0-Fundamental of Compression
17 pages
Data Compression Report
No ratings yet
Data Compression Report
12 pages
Designing and Building Parallel Programs
No ratings yet
Designing and Building Parallel Programs
371 pages
Data Compression
No ratings yet
Data Compression
21 pages
Data Compression
No ratings yet
Data Compression
29 pages
MEGA65-Book Draft PDF
No ratings yet
MEGA65-Book Draft PDF
822 pages
Image Compression: CS474/674 - Prof. Bebis
No ratings yet
Image Compression: CS474/674 - Prof. Bebis
110 pages
A Jump Start To Opencl: March 15, 2009 Cis 565/665 - Gpu Computing and Architecture
No ratings yet
A Jump Start To Opencl: March 15, 2009 Cis 565/665 - Gpu Computing and Architecture
74 pages
Egoless Programming
No ratings yet
Egoless Programming
2 pages
Study of LZ77 and LZ78 Data Compression
No ratings yet
Study of LZ77 and LZ78 Data Compression
5 pages
Ogre Tutorial
100% (1)
Ogre Tutorial
57 pages
Vulkan Tutorial PDF
No ratings yet
Vulkan Tutorial PDF
293 pages
Lemp El Ziv Report
No ratings yet
Lemp El Ziv Report
17 pages
Compression
No ratings yet
Compression
46 pages
Text Compression
No ratings yet
Text Compression
25 pages
Direct2D Succinctly PDF
No ratings yet
Direct2D Succinctly PDF
187 pages
Vulkan Tutorial
No ratings yet
Vulkan Tutorial
239 pages
GPU Programming Using openCL
No ratings yet
GPU Programming Using openCL
13 pages
1.3 Data Storage and File Compression
No ratings yet
1.3 Data Storage and File Compression
6 pages
System Design With SystemC
No ratings yet
System Design With SystemC
236 pages
Image Compression
No ratings yet
Image Compression
111 pages
Tetris Tutorial in C++ Platform Independent Focused in Game Logic For Beginners - Javilop
100% (2)
Tetris Tutorial in C++ Platform Independent Focused in Game Logic For Beginners - Javilop
44 pages
Amiga World Tech Journal Vol 01-01-1991 Apr May
No ratings yet
Amiga World Tech Journal Vol 01-01-1991 Apr May
56 pages
Lempel-Ziv-Welch (LZW) Compression Algorithm
No ratings yet
Lempel-Ziv-Welch (LZW) Compression Algorithm
22 pages
8080 z80 Assembly Language Techniques For Improved Programming
No ratings yet
8080 z80 Assembly Language Techniques For Improved Programming
332 pages
Compiling With Continuations
No ratings yet
Compiling With Continuations
270 pages
The Architecture of Massive Multiplayer Online Games: Sladjan Bogojevic, Mohsen Kazemzadeh
No ratings yet
The Architecture of Massive Multiplayer Online Games: Sladjan Bogojevic, Mohsen Kazemzadeh
107 pages
Amiga World Tech Journal Vol 02-01-1992 Feb
No ratings yet
Amiga World Tech Journal Vol 02-01-1992 Feb
68 pages
Vulkan
No ratings yet
Vulkan
77 pages
GBA Programming Manual v1.22
No ratings yet
GBA Programming Manual v1.22
172 pages
SIGGRAPH 2015 Remedy Notes PDF
No ratings yet
SIGGRAPH 2015 Remedy Notes PDF
164 pages
Atari Harddisk Quick Faq
No ratings yet
Atari Harddisk Quick Faq
6 pages
Commodore Magazine Vol-08-N04 1987 Apr
No ratings yet
Commodore Magazine Vol-08-N04 1987 Apr
132 pages
Game Boy Advance - Midi Interface (Hack Part 2)
100% (5)
Game Boy Advance - Midi Interface (Hack Part 2)
7 pages
Amiga World Tech Journal Vol 01-04-1991 Oct
No ratings yet
Amiga World Tech Journal Vol 01-04-1991 Oct
74 pages
Prince of Persia
No ratings yet
Prince of Persia
11 pages
Game Boy CPU Manual
No ratings yet
Game Boy CPU Manual
139 pages
Amiga World Tech Journal Vol 01-02-1991 Jun Jul
No ratings yet
Amiga World Tech Journal Vol 01-02-1991 Jun Jul
56 pages
GBProject PDF
No ratings yet
GBProject PDF
28 pages
0864 Introduction To Programming Using Fortran 9520032008 PDF
No ratings yet
0864 Introduction To Programming Using Fortran 9520032008 PDF
237 pages
Data Compression
No ratings yet
Data Compression
201 pages
Co3325 ch1-4
No ratings yet
Co3325 ch1-4
47 pages
Data Compression
No ratings yet
Data Compression
46 pages
Gray Hat Hacking the Ethical Hacker's
From Everand
Gray Hat Hacking the Ethical Hacker's
Çağatay Şanlı
5/5 (1)
GED Test 5-Hour Quick Prep For Dummies
From Everand
GED Test 5-Hour Quick Prep For Dummies
Tim Collins
No ratings yet
Nailing Jelly to the Wall: Defining and Providing Technical Assistance in Early Childhood Education
From Everand
Nailing Jelly to the Wall: Defining and Providing Technical Assistance in Early Childhood Education
Nancy P Alexander
5/5 (1)
Programming the Photon: Getting Started with the Internet of Things
From Everand
Programming the Photon: Getting Started with the Internet of Things
Christopher Rush
5/5 (1)
Programming the Intel Galileo: Getting Started with the Arduino -Compatible Development Board
From Everand
Programming the Intel Galileo: Getting Started with the Arduino -Compatible Development Board
Christopher Rush
5/5 (1)
GRE 5-Hour Quick Prep For Dummies
From Everand
GRE 5-Hour Quick Prep For Dummies
Ron Woldoff
No ratings yet
Teardowns: Learn How Electronics Work by Taking Them Apart
From Everand
Teardowns: Learn How Electronics Work by Taking Them Apart
Bryan Bergeron
No ratings yet
Programming Arduino: Getting Started with Sketches
From Everand
Programming Arduino: Getting Started with Sketches
Simon Monk
3.5/5 (5)
Project Location Map PDF
No ratings yet
Project Location Map PDF
1 page
Microsoft Dynamics 365 For Finance and Operations Keyboard Shortcuts PDF
No ratings yet
Microsoft Dynamics 365 For Finance and Operations Keyboard Shortcuts PDF
1 page
Web
No ratings yet
Web
187 pages
ALS1
No ratings yet
ALS1
3 pages
Game Api
No ratings yet
Game Api
16 pages
Datacolor Portable Spectrophotometers
No ratings yet
Datacolor Portable Spectrophotometers
2 pages
Testdisk Documentation: Release 7.1
No ratings yet
Testdisk Documentation: Release 7.1
69 pages
Epson L310-L360
No ratings yet
Epson L310-L360
2 pages
Manuale Di Istruzioni: Instruction Manual
No ratings yet
Manuale Di Istruzioni: Instruction Manual
33 pages
NSE 2 - Sandbox
60% (5)
NSE 2 - Sandbox
2 pages
CS Q (C)
No ratings yet
CS Q (C)
10 pages
Unix Shell Scripting
100% (1)
Unix Shell Scripting
46 pages
PAL4 UserManual
No ratings yet
PAL4 UserManual
6 pages
Unit 5 Software Engineering
No ratings yet
Unit 5 Software Engineering
18 pages
Zos Admin Freshers
100% (1)
Zos Admin Freshers
6 pages
Sanet - CD Computeractive Issue 505 518 July 2017 PDF
No ratings yet
Sanet - CD Computeractive Issue 505 518 July 2017 PDF
76 pages
ORACLE 10g Installation
No ratings yet
ORACLE 10g Installation
6 pages
Tech Spec ASN 18 GHZ
No ratings yet
Tech Spec ASN 18 GHZ
12 pages
CnWave 60 GHZ User Guide 1.2
No ratings yet
CnWave 60 GHZ User Guide 1.2
208 pages
How To Use Clonezilla
No ratings yet
How To Use Clonezilla
54 pages
DS Lab 5 - Array Based Stack Implementation
No ratings yet
DS Lab 5 - Array Based Stack Implementation
14 pages
V2216 Series: Network Video Management Software
No ratings yet
V2216 Series: Network Video Management Software
4 pages
Inventory Management: Exercise 5.1
No ratings yet
Inventory Management: Exercise 5.1
11 pages
2019 Winter Model Answer Paper (Msbte Study Resources)
No ratings yet
2019 Winter Model Answer Paper (Msbte Study Resources)
23 pages
Ip PDF
No ratings yet
Ip PDF
25 pages
TCP Concurent Server
No ratings yet
TCP Concurent Server
13 pages
Abstract 1
No ratings yet
Abstract 1
2 pages
Installation Guide For Scopia Elite 6000 For Aura Collaboration Suite Version 82
No ratings yet
Installation Guide For Scopia Elite 6000 For Aura Collaboration Suite Version 82
60 pages
Literature Review of Student Record Management System
No ratings yet
Literature Review of Student Record Management System
7 pages