Data Compression

introduction basic coding schemes an application entropy LZW codes

References: Algorithms 2nd edition, Chapter 22
http://www.cs.princeton.edu/introalgsds/65compression
1

 introduction basic coding schemes an application entropy LZW codes

Data Compression Compression reduces the size of a file: To save space when storing it. To save time when transmitting it. Most files have lots of redundancy.

Who needs compression? Moore's law: # transistors on a chip doubles every 18-24 months. Parkinson's law: data expands to fill space available. Text, images, sound, video,

All of the books in the world contain no more information than is broadcast as video in a single large American city in a single year. Not all bits have equal value. -Carl Sagan

Basic concepts ancient (1950s), best technology recently developed.

3

Applications Generic file compression. Files: GZIP, BZIP, BOA. Archivers: PKZIP. File systems: NTFS.

Multimedia. Images: GIF, JPEG. Sound: MP3. Video: MPEG, DivX, HDTV. Communication. ITU-T T4 Group 3 Fax. V.42bis modem. Databases. Google.

Encoding and decoding

uses fewer bits (you hope) Message. Binary data M we want to compress. Encode. Generate a "compressed" representation C(M). Decode. Reconstruct original message or some approximation M'.

Encoder

C(M)

Decoder

M'

Compression ratio. Bits in C(M) / bits in M. Lossless. M = M', 50-75% or lower. Ex. Natural language, source code, executables. Lossy. M M', 10% or lower. Ex. Images, sound, video.
this lecture

"Poetry is the art of lossy data compression."

Food for thought Data compression has been omnipresent since antiquity, Number systems. Natural languages. Mathematical notation.

has played a central role in communications technology, Braille. Morse code. Telephone system. and is part of modern life. zip. MP3. MPEG. What role will it play in the future?
Ex: If memory is to be cheap and ubiquitous, why are we doing lossy compression for music and movies??
6

 introduction basic coding schemes an application entropy LZW codes

Fixed length encoding

Use same number of bits for each symbol. k-bit code supports 2 different symbols
k

Ex. 7-bit ASCII

char decimal code

NUL ... a b c d ... ~

0 ... 97 98 99 100 ... 126 127

0 1100001 1100010 1100011 1100100 1111110 1111111 d a b r

this lecture: special code for end-of-message

1100001 1100010 1110010 1100001 1100011 1100001 1100100 1100001 1100010 1110010 1100001 1111111

12 symbols 7 bits per symbol = 84 bits in code

8

Fixed length encoding

Use same number of bits for each symbol. k-bit code supports 2 different symbols
k

Ex. 3-bit custom code

char code

a b c d r ! a b r a c a d

X000 X001 X010 X011 X100 X111 a b r a !

12 symbols 3 bits 36 bits in code

X000 X001 X100 X000 X010 X000 X011 X000 X001 X100 X000 X111

Important detail: decoder needs to know the code!

Fixed length encoding: general scheme

count number of different symbols. lg M bits suffice to support M different symbols
Ex. genomic sequences 4 different codons 2 bits suffice

char

code

a c t g

X00 X01 X10 X11

g
11

2N bits to encode genome with N codons

t
10

a
X00X

t
10

g
11

a
X00X

X00X X01X

X00X X01X X00X

Amazing but true: initial databases in 1990s did not use such a code!
Decoder needs to know the code can amortize over large number of files with the same code in general, can encode an N-char file with N lg M + 16 lg M bits

10

Variable-length encoding Use different number of bits to encode different characters. Ex. Morse code. Issue: ambiguity.
SOS ? IAMIE ? EEWNI ? V7O ?

11

Variable-length encoding Use different number of bits to encode different characters. Q. How do we avoid ambiguity? A1. Append special stop symbol to each codeword. A2. Ensure that no encoding is a prefix of another. Ex. custom prefix-free code
char code

S E I V

prefix of V prefix of I, S prefix of S

a b c d r !

0 111 1010 100 110 1011

28 bits in code

a b r a c a d a b r a ! 0 1 1 1 1 1 0 0 1 0 1 0 0 1 0 0 0 1 1 1 1 1 0 0 1 0 1 1

Note 1: fixed-length codes are prefix-free Note 2: can amortize cost of including the code over similar messages

12

Prefix-free code: Encoding and Decoding How to represent? Use a binary trie. Symbols are stored in leaves. Encoding is path to leaf.

Encoding. Method 1: start at leaf; follow path up to the root, and print bits in reverse order. Method 2: create ST of symbol-encoding pairs. Decoding. Start at root of tree. Go left if bit is 0; go right if 1. If leaf node, print symbol and return to root.

1 0 1

d c !

char
a b c d r !

encoding
0 111 1010 100 110 1011

13

Providing the code How to transmit the trie? send preorder traversal of trie. we use * as sentinel for internal nodes [ what if no sentinel is available? ] send number of characters to decode. send bits (packed 8 to the byte).

1 0 1

d c ! *a**d*c!*rb 12 0111110010100100011111001011

preorder traversal # chars to decode the message bits

If message is long, overhead of transmitting trie is small.

char
a b c d

encoding
0 111 1010 100 110 1011
14

r !

0 1 1 1 1 1 0 0 1 0 1 0 0 1 0 0 0 1 1 1 1 1 0 0 1 0 1 1

Prefix-free decoding implementation

public class PrefixFreeDecoder { private Node root = new Node(); private class Node { char ch; Node left, right; Node() { ch = StdIn.readChar(); if (ch == '*') { left = new Node(); right = new Node(); } } boolean isInternal() { } } public void decode() { /* See next slide. */ }

1 0 1

d c
build tree from preorder traversal

r !

*a**d*c!*rb

15

Prefix-free decoding iImplementation

public void decode() { int N = StdIn.readInt(); for (int i = 0; i < N; i++) { Node x = root; while (x.isInternal()) { use bits, not chars char bit = StdIn.readChar(); in actual applications if (bit == '0') x = x.left; else if (bit == '1') x = x.right; } System.out.print(x.ch); } }
more code.txt 12 0111110010100100011111001011 % java PrefixFreeDecoder < code.txt abacadabra!

1 0 1

d c !

16

Introduction to compression: summary Variable-length codes can provide better compression than fixed-length
a b r a c a d a b r a !

1100001 1100010 1110010 1100001 1100011 1100001 1100100 1100001 1100010 1110010 1100001 1111111

X000 X001 X100 X000 X010 X000 X011 X000 X001 X100 X000 X111

a b

r a

a d a b

r a

0111110010100100011111001011

Every trie defines a variable-length code

Q. What is the best variable length code for a given message?

17

Huffman coding Q. What is the best variable length code for a given message? A. Huffman code. [David Huffman, 1950]

To compute Huffman code: count frequency ps for each symbol s in message.

start with one node corresponding to each symbol s (with weight p ). repeat until single trie formed:
s

select two tries with min weight p1 and p2 merge into single trie with weight p1 + p2

Applications. JPEG, MP3, MPEG, PKZIP, GZIP,

David Huffman
18

Huffman coding example

a
1

b
1

r
1

c
2

a
2

d
5

!
5 7 4 2 2 1 1 2

a c ! d
1

r
2 1

a
2 2 5 1

d
1

r c !

d c !

3 1 2 1 1

12

b d c !

a
5 7 4 2 2 1 1 2

a
3 1

d
3 1 2 2 1 1 4 2 5

r !

a d c ! r b

19

Huffman trie construction code

int[] freq = new int[128]; for (int i = 0; i < input.length(); i++) { freq[input.charAt(i)]++; } MinPQ<Node> pq = new MinPQ<Node>(); for (int i = 0; i < 128; i++) if (freq[i] > 0) pq.insert(new Node((char) i, freq[i], null, null)); while (pq.size() > 1) { Node x = pq.delMin(); Node y = pq.delMin(); Node parent = new Node('*', x.freq + y.freq, x, y); pq.insert(parent); } two subtrees root = pq.delMin();
internal node marker total frequency

tabulate frequencies

initialize PQ

merge trees

20

Huffman encoding summary Theorem. Huffman coding is an optimal prefix-free code.

no prefix-free code uses fewer bits

Implementation. pass 1: tabulate symbol frequencies and build trie pass 2: encode file by traversing trie or lookup table.

Running time. Use binary heap O(M + N log N).

output bits distinct symbols

Can we do better? [stay tuned]

21

 introduction basic coding schemes an application entropy LZW codes

22

An application: compress a bitmap Typical black-and-white-scanned image 300 pixels/inch 8.5 by 11 inches 300*8.5*300*11 = 8.415 million bits Bits are mostly white

Typical amount of text on a page: 40 lines * 75 chars per line = 3000 chars

23

Natural encoding of a bitmap one bit per pixel

000000000000000000000000000011111111111111000000000 000000000000000000000000001111111111111111110000000 000000000000000000000001111111111111111111111110000 000000000000000000000011111111111111111111111111000 000000000000000000001111111111111111111111111111110 000000000000000000011111110000000000000000001111111 000000000000000000011111000000000000000000000011111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000001111000000000000000000000001110 000000000000000000000011100000000000000000000111000 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011000000000000000000000000000000000000000000000011

19-by-51 raster of letter 'q' lying on its side

24

Run-length encoding of a bitmap

to encode number of bits per line

natural encoding. (19 51) + 6 = 975 bits. run-length encoding. (63 6) + 6 = 384 bits.
63 6-bit run lengths 000000000000000000000000000011111111111111000000000 000000000000000000000000001111111111111111110000000 000000000000000000000001111111111111111111111110000 000000000000000000000011111111111111111111111111000 000000000000000000001111111111111111111111111111110 000000000000000000011111110000000000000000001111111 000000000000000000011111000000000000000000000011111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000011100000000000000000000000000111 000000000000000000001111000000000000000000000001110 000000000000000000000011100000000000000000000111000 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011111111111111111111111111111111111111111111111111 011000000000000000000000000000000000000000000000011 51 28 14 9 26 18 7 23 24 4 22 26 3 20 30 1 19 7 18 7 19 5 22 5 19 3 26 3 19 3 26 3 19 3 26 3 19 3 26 3 20 4 23 3 1 22 3 20 3 3 1 50 1 50 1 50 1 50 1 50 1 2 46 2

19-by-51 raster of letter 'q' lying on its side

RLE
25

Run-length encoding

Exploit long runs of repeated characters. Bitmaps: runs alternate between 0 and 1; just output run lengths. Issue: how to encode run lengths (!)
10: 2 01: 1 001001001001001 2121212121 10011001100110011001

15 bits

20 bits

Does not compress when runs are short.

Runs are long in typical applications (such as black-and-white bitmaps).

26

Run-length encoding and Huffman codes in the wild ITU-T T4 Group 3 Fax for black-and-white bitmap images (~1980) up to 1728 pixels per line typically mostly white.

one for white and one for black Step 1. Use run-length encoding. Step 2. Encode run lengths using two Huffman codes. run
0 1 2

194

white
00110101 000111 0111 1000 00110100 11011 10010 010011011

black
0000110111 010 11 10 000001100111 0000001111 000011001000 0000001100101
27

3W

1B

2W

2B

194W

1000

010 0111

11

010111 0111 192 + 2

63 64+ 128+

Huffman codes built from frequencies in huge sample

1728+

BW bitmap compression: another approach Fax machine (~1980) slow scanner produces lines in sequential order compress to save time (reduce number of bits to send)

Electronic documents (~2000) high-resolution scanners produce huge files compress to save space (reduce number of bits to save) Idea: use OCR to get back to ASCII (!) use Huffman on ASCII string (!) Ex. Typical page 40 lines, 75 chars/line ~ 3000 chars compress to ~ 2000 chars with Huffman code reduce file size by a factor of 500 (! ?) Bottom line: Any extra information about file can yield dramatic gains
28

 introduction basic coding schemes an application entropy LZW codes

29

What data can be compressed? US Patent 5,533,051 on "Methods for Data Compression", which is capable of compression all files.

Slashdot reports of the Zero Space Tuner and BinaryAccelerator.

"ZeoSync has announced a breakthrough in data compression that allows for 100:1 lossless compression of random data. If this is true, our bandwidth problems just got a lot smaller."

30

Perpetual Motion Machines Universal data compression algorithms are the analog of perpetual motion machines.

Closed-cycle mill by Robert Fludd, 1618

Gravity engine by Bob Schadewald

Reference: Museum of Unworkable Devices by Donald E. Simanek http://www.lhup.edu/~dsimanek/museum/unwork.htm

31

What data can be compressed? Theorem. Impossible to losslessly compress all files. Pf 1. consider all 1,000 bit messages. 21000 possible messages. only 2999 + 2998 + + 1 can be encoded with 999 bits. only 1 in 2499 can be encoded with 500 bits!

Pf 2 (by contradiction). given a file M, compress it to get a smaller file M1. compress that file to get a still smaller file M2. continue until reaching file size 0. implication: all files can be compressed with 0 bits!
Practical test for any compression algorithm: given a file M, compress it to get a (smaller, you hope) file M1 compress that file to get a still smaller file M2. continue until file size does not decrease
32

A difficult file to compress

One million pseudo-random characters (a p)

fclkkacifobjofmkgdcoiicnfmcpcjfccabckjamolnihkbgobcjbngjiceeelpfgcjiihppenefllhglfemdemgahlbpi ggmllmnefnhjelmgjncjcidlhkglhceninidmmgnobkeglpnadanfbecoonbiehglmpnhkkamdffpacjmgojmcaabpcjce cplfbgamlidceklhfkkmioljdnoaagiheiapaimlcnlljniggpeanbmojgkccogpmkmoifioeikefjidbadgdcepnhdpfj aeeapdjeofklpdeghidbgcaiemajllhnndigeihbebifemacfadnknhlbgincpmimdogimgeeomgeljfjgklkdgnhafoho npjbmlkapddhmepdnckeajebmeknmeejnmenbmnnfefdbhpmigbbjknjmobimamjjaaaffhlhiggaljbaijnebidpaeigd goghcihodnlhahllhhoojdfacnhadhgkfahmeaebccacgeojgikcoapknlomfignanedmajinlompjoaifiaejbcjcdibp kofcbmjiobbpdhfilfajkhfmppcngdneeinpnfafaeladbhhifechinknpdnplamackphekokigpddmmjnbngklhibohdf eaggmclllmdhafkldmimdbplggbbejkcmhlkjocjjlcngckfpfakmnpiaanffdjdlleiniilaenbnikgfnjfcophbgkhdg mfpoehfmkbpiaignphogbkelphobonmfghpdgmkfedkfkchceeldkcofaldinljjcgafimaanelmfkokcjekefkbmegcgj ifjcpjppnabldjoaafpbdafifgcoibbcmoffbbgigmngefpkmbhbghlbdjngenldhgnfbdlcmjdmoflhcogfjoldfjpaok epndejmnbiealkaofifekdjkgedgdlgbioacflfjlafbcaemgpjlagbdgilhcfdcamhfmppfgohjphlmhegjechgdpkklj pndphfcnnganmbmnggpphnckbieknjhilafkegboilajdppcodpeoddldjfcpialoalfeomjbphkmhnpdmcpgkgeaohfdm cnegmibjkajcdcpjcpgjminhhakihfgiiachfepffnilcooiciepoapmdjniimfbolchkibkbmhbkgconimkdchahcnhap fdkiapikencegcjapkjkfljgdlmgncpbakhjidapbldcgeekkjaoihbnbigmhboengpmedliofgioofdcphelapijcegej gcldcfodikalehbccpbbcfakkblmoobdmdgdkafbbkjnidoikfakjclbchambcpaepfeinmenmpoodadoecbgbmfkkeabi laoeoggghoekamaibhjibefmoppbhfbhffapjnodlofeihmjahmeipejlfhloefgmjhjnlomapjakhhjpncomippeanbik khekpcfgbgkmklipfbiikdkdcbolofhelipbkbjmjfoempccneaebklibmcaddlmjdcajpmhhaeedbbfpjafcndianlfcj mmbfncpdcccodeldhmnbdjmeajmboclkggojghlohlbhgjkhkmclohkgjamfmcchkchmiadjgjhjehflcbklfifackbecg joggpbkhlcmfhipflhmnmifpjmcoldbeghpcekhgmnahijpabnomnokldjcpppbcpgcjofngmbdcpeeeiiiclmbbmfjkhl anckidhmbeanmlabncnccpbhoafajjicnfeenppoekmlddholnbdjapbfcajblbooiaepfmmeoafedflmdcbaodgeahimc gpcammjljoebpfmghogfckgmomecdipmodbcempidfnlcggpgbffoncajpncomalgoiikeolmigliikjkolgolfkdgiijj iooiokdihjbbofiooibakadjnedlodeeiijkliicnioimablfdpjiafcfineecbafaamheiipegegibioocmlmhjekfikf effmddhoakllnifdhckmbonbchfhhclecjamjildonjjdpifngbojianpljahpkindkdoanlldcbmlmhjfomifhmncikol jjhebidjdphpdepibfgdonjljfgifimniipogockpidamnkcpipglafmlmoacjibognbplejnikdoefccdpfkomkimffgj gielocdemnblimfmbkfbhkelkpfoheokfofochbmifleecbglmnfbnfncjmefnihdcoeiefllemnohlfdcmbdfebdmbeeb balggfbajdamplphdgiimehglpikbipnkkecekhilchhhfaeafbbfdmcjojfhpponglkfdmhjpcieofcnjgkpibcbiblfp njlejkcppbhopohdghljlcokhdoahfmlglbdkliajbmnkkfcoklhlelhjhoiginaimgcabcfebmjdnbfhohkjphnklcbhc jpgbadakoecbkjcaebbanhnfhpnfkfbfpohmnkligpgfkjadomdjjnhlnfailfpcmnololdjekeolhdkebiffebajjpclg hllmemegncknmkkeoogilijmmkomllbkkabelmodcohdhppdakbelmlejdnmbfmcjdebefnjihnejmnogeeafldabjcgfo aehldcmkbnbafpciefhlopicifadbppgmfngecjhefnkbjmliodhelhicnfoongngemddepchkokdjafegnpgledakmbcp cmkckhbffeihpkajginfhdolfnlgnadefamlfocdibhfkiaofeegppcjilndepleihkpkkgkphbnkggjiaolnolbjpobjd cehglelckbhjilafccfipgebpc....

33

A difficult file to compress

public class Rand { public static void main(String[] args) { for (int i = 0; i < 1000000; i++) { char c = 'a'; c += (char) (Math.random() * 16); System.out.print(c); } } }
231 bytes, but output is hard to compress (assume random seed is fixed)

% % % % %

javac Rand.java java Rand > temp.txt compress c temp.txt > temp.Z gzip c temp.txt > temp.gz bzip2 c temp.txt > temp.bz2

% ls l 231 1000000 576861 570872 499329

Rand.java temp.txt temp.Z temp.gz temp.bz2

resulting file sizes (bytes)

34

Information theory Intrinsic difficulty of compression. Short program generates large data file. Optimal compression algorithm has to discover program! Undecidable problem.

Q. How do we know if our algorithm is doing well? A. Want lower bound on # bits required by any compression scheme.

35

Language model Q. How do compression algorithms work? A. They exploit statistical biases of input messages. ex: white patches occur in typical images. ex: ord Princeton occurs more frequently than Yale.

Basis of compression: probability. Formulate probabilistic model to predict symbols. simple: character counts, repeated strings complex: models of a human face Use model to encode message. Use same model to decode message. Ex. Order 0 Markov model R symbols generated independently at random probability of occurrence of i th symbol: pi (fixed).

36

Entropy A measure of information. [Shannon, 1948] H(M) = p0/lg p0 + p1/lg p1 + p2/lg p2 + ... + pR-1/lg pR-1

information content of symbol s is proportional to 1/lg p(s). weighted average of information content over all symbols. interface between coding and model.
2

Ex. 4 binary models (R = 2)

p0 1 2 3 4 1/2 0.900 0.990 1 p1 1/2 0.100 0.010 0 H(M) 1 0.469 0.0808 0
Claude Shannon

Ex. fair die (R = 6)

p(1) 1/6 p(2) 1/6 p(3) 1/6 p(4) 1/6 p(5) 1/6 p(6) 1/6 H(M) 2.585
37

Entropy and compression Theorem. [Shannon, 1948] If data source is an order 0 Markov model, any compression scheme must use H(M) bits per symbol on average.

Cornerstone result of information theory. Ex: to transmit results of fair die, need 2.58 bits per roll.
Theorem. [Huffman, 1952] If data source is an order 0 Markov model, Huffman code uses H(M) + 1 bits per symbol on average.

Q. Is there any hope of doing better than Huffman coding? A1. Yes. Huffman wastes up to 1 bit per symbol. if H(M) is close to 0, this difference matters can do better with "arithmetic coding" A2. Yes. Source may not be order 0 Markov model.
38

Entropy of the English Language Q. How much redundancy is in the English language?

"... randomising letters in the middle of words [has] little or no effect on the ability of skilled readers to understand the text. This is easy to denmtrasote. In a pubiltacion of New Scnieitst you could ramdinose all the letetrs, keipeng the first two and last two the same, and reibadailty would hadrly be aftcfeed. My ansaylis did not come to much beucase the thoery at the time was for shape and senqeuce retigcionon. Saberi's work sugsegts we may have some pofrweul palrlael prsooscers at work. The resaon for this is suerly that idnetiyfing coentnt by paarllel prseocsing speeds up regnicoiton. We only need the first and last two letetrs to spot chganes in meniang."

A. Quite a bit.

39

Entropy of the English Language Q. How much information is in each character of the English language? Q. How can we measure it?
model = English text

A. [Shannon's 1951 experiment] Asked subjects to predict next character given previous text. The number of guesses required for right answer:

# of guesses Fraction

1 0.79

2 0.08

3 0.03

4 0.02

5 0.02

6 0.05

Shannon's estimate: about 1 bit per char

[ 0.6 - 1.3 ].

Compression less than 1 bit/char for English ? If not, keep trying!

40

 introduction basic coding schemes an application entropy LZW codes

41

Statistical Methods Static model. Same model for all texts. Fast. Not optimal: different texts have different statistical properties. Ex: ASCII, Morse code.

Dynamic model. Generate model based on text. Preliminary pass needed to generate model. Must transmit the model. Ex: Huffman code. Adaptive model. Progressively learn and update model as you read text. More accurate modeling produces better compression. Decoding must start from beginning. Ex: LZW.

42

LZW Algorithm Lempel-Ziv-Welch. [variant of LZ78] Create ST associating a fixed-length codeword with some previous substring. When input matches string in ST, output associated codeword. length of strings in ST grows, hence compression.

To send (encode) M. Find longest string s in ST that is a prefix of unsent part of M Send codeword associated with s. Add s x to ST, where x is next char in M.

Ex. ST: a, aa, ab, aba, abb, abaa, abaab, abaaa, unsent part of M: abaababbb s = abaab, x = a. Output integer associated with s; insert abaaba into ST.

43

LZW encoding example

input code add to ST

a b r a c a d a b r a c a d a b r a STOP

97 98 114 97 99 97 100

ab br ra ac ca

input: 7-bit ASCII output: 8-bit codewords

ASCII
key value 0 key ab br ... ra ac a 97 98 99 100 ... r 114 ca ad da abr rac cad dab bra

ST
value 128 129 130 131 132 133 134 135 136 137 138 139 ... STOP 255

ad da

b c d

128

abr

130

rac ... 127

132

cad it_

134

dab

129 97 255

bra

To send (encode) M. Find longest string s in ST that is a prefix of unsent part of M Send integer associated with s. Add s x to ST, where x is next char in M.

44

LZW encoding example

input code

a b r a c a d

97 98 114 97 99 97 100

input:

7-bit ASCII 19 chars 133 bits

a b r a c a d a b r a STOP 129 97 255 134 132 130 128

output: 8-bit codewords 14 chars 112 bits

Key point: no need to send ST (!)

45

LZW encode ST implementation Q. How to do longest prefix match? A. Use a trie for the ST Encode. lookup string suffix in trie. output ST index at bottom. add new node to bottom of trie.
ASCII
key value 0 key ab br ... ra ac a b c d 97 98 99 100 ...
97 98 99 100 114

ST
value 128 129 130 131 132 133 134 135 136 137 138 139 ... STOP 255

ca ad da abr rac cad dab bra

114

a
128 131

b
133 129

c
132

d
134

r
130

... 127

b
135

r
139

a
137

a
138 136

Note that all substrings are in ST

46

LZW encoder: Java implementation

public class LZWEncoder { public static void main(String[] args) { LookAheadIn in = new LookAheadIn(); LZWst st = new LZWst(); while (!in.isEmpty()) { int codeword = st.getput(in); StdOut.println(codeword); } } }
postprocess to encode in binary

input stream with lookahead

specialized TST

encode text and build TST

Use specialized TST initialized with ASCII chars and codes getput() method returns code of longest prefix s and adds s + next char to symbol table

Need input stream with backup [stay tuned]

47

LZW encoder: Java implementation (TST scaffolding)

public class LZWst { private int i; private int codeword; private Node[] roots;

next codeword to assign codeword to return array of TSTs

public LZWst() { roots = new Node[128]; for (i = 0; i < 128; i++) roots[i] = new Node((char) i, i); } private class Node { Node(char c, int codeword) { this.c = c; this.codeword = codeword; char c; Node left, mid, right; int codeword; } public int getput(LookAheadIn in) // See next slide. }

initialize with ASCII

standard node code

48

LZW encoder: Java implementation (TST search/insert)

public int getput(LookAheadIn in) { char c = in.readChar(); if (c == '!') return 255; roots[c] = getput(c, roots[c], in); in.backup(); longest prefix return codeword; codeword } public Node getput(char c, Node x, LookAheadIn in) { if (x == null) { x = new Node(c, i++); return x; } if (c < x.c) x.left = getput(c, x.left, in); else if (c > x.c) x.right = getput(c, x.right, in); else { char next = in.readChar(); codeword = x.codeword; x.mid = getput(next, x.mid, in); } return x; }
check for codeword overflow omitted

caution: tricky recursive code

recursive search and insert

49

LZW encoder: Java implementation (input stream with lookahead)

public class LookAheadIn { In in = new In(); char last; boolean backup = false; public void backup() { backup = true; } public char readChar() { if (!backup) { last = in.readChar(); backup = false; return last; }

public boolean isEmpty() { return !backup && in.isEmpty(); }

Provides input stream with one-character lookahead. backup() call means that last readChar() call was lookahead.

50

LZW Algorithm Lempel-Ziv-Welch. [variant of LZ78] Create ST and associate an integer with each useful string. When input matches string in ST, output associated integer. length of strings in ST grows, hence compression. decode by rebuilding ST from code

To send (encode) M. Find longest string s in ST that is a prefix of unsent part of M Send integer associated with s. Add s x to ST, where x is next char in M.

To decode received message to M. Let s be ST entry associated with received integer Add s to M. Add p x to ST, where x is first char in s, p is previous value of s.
51

LZW decoding example

codeword output add to ST

role of keys and values switched

key value key 128 129 ... 130 131 97 98 99 a b c d 132 133 134 135 136 r 137 138 ... 139 ... 127 255 value ab br ra ac ca ad da abr rac cad dab bra

97 98 114 97 99 97 100 128

a b r a c a d a b da ab br ra ac ca ad

100 ... 114 abr

Use an array to implement ST

130

r a

132

c a rac it_ cad

134

d a

129

b r dab bra

97 255

a STOP

To decode received message to M. Let s be ST entry associated with received integer Add s to M. Add p x to ST, where x is first char in s, p is previous value of s.

52

LZW decoder: Java implementation

public class LZWDecoder { public static void main(String[] args) { String[] st = new String[256]; int i; for (i = 0; i < 128; i++) { st[i] = Character.toString((char) i); st[255] = "!";

initialize ST with ASCII

}
decode text and build ST

String prev = ""; to decode while (!StdIn.isEmpty()) from binary { int codeword = StdIn.readInt(); String s; if (codeword == i) // Tricky situation! s = prev + prev.charAt(0); else s = st[codeword]; StdOut.print(s); if (prev.length() > 0) { st[i++] = prev + s.charAt(0); } prev = s; } StdOut.println(); } }

preprocess

Ex: ababababab

53

LZW decoding example (tricky situation)

input

code

add to ST

codeword

output

add to ST

a b a b a b a b STOP

97 98

ab ba

key 128 129

value ab ba aba

97 98 128

a b a b ba ab

128

aba

130 131

130 abab 130 abab ... 255 255 98 255

a b a b STOP aba

needed before added to ST! To send (encode) M. Find longest prefix Send integer associated with s. Add s x to ST, where x is next char in M.

To decode received message to M. Let s be ST entry for integer Add s to M. Add p x to ST where x is first char in s p is previous value of s.

54

LZW implementation details How big to make ST? how long is message? whole message similar model? ... [many variations have been developed]

What to do when ST fills up? throw away and start over. GIF throw away when not effective. Unix compress ... [many other variations] Why not put longer substrings in ST? ... [many variations have been developed]

55

LZW in the real world Lempel-Ziv and friends. LZ77. LZ77 not patented widely used in open source LZ78. LZW patent #4,558,302 expired in US on June 20, 2003 some versions copyrighted LZW. Deflate = LZ77 variant + Huffman.

PNG: LZ77. Winzip, gzip, jar: deflate. Unix compress: LZW. Pkzip: LZW + Shannon-Fano. GIF, TIFF, V.42bis modem: LZW. Google: zlib which is based on deflate.
never expands a file

56

Lossless compression ratio benchmarks Calgary corpus: standard data compression benchmark
Year 1967 1950 1977 1984 1987 1987 1987 1987 1988 1988 1994 1995 1997 1999 Scheme ASCII Huffman LZ77 LZMW LZH Move-to-front LZB Gzip PPMC SAKDC PPM Burrows-Wheeler BOA RK Bits / char 7.00 4.70 3.94 3.32 3.30 3.24 3.18 2.71 2.48 2.47 2.34 2.29 1.99 1.89
57

Entropy Char by char 8 chars at a time Asymptotic

Bits/char 4.5 2.4 1.3

next assignment

Data compression summary Lossless compression. Represent fixed length symbols with variable length codes. [Huffman] Represent variable length symbols with fixed length codes. [LZW]

Lossy compression. [not covered in this course] JPEG, MPEG, MP3. FFT, wavelets, fractals, SVD, Limits on compression. Shannon entropy. Theoretical limits closely match what we can achieve in practice. Practical compression: Use extra knowledge whenever possible.
Butch: I dont mean to be a sore loser, but when its done, if Im dead, kill him. Sundance: Love to. Butch: No, no, not yet. Not until me and Harvey get the rules straightened out. Harvey: Rules? In a knife fight? No rules. Butch: Well, if there aint going to be any rules, lets get the fight started...
58