Compression

Lossless Compression
Lossy image compression

Image Compression
EE 604

Govind Sharma

Department of Electrical Engineering

Indian Institute of Technology, Kanpur

3/4/2019

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Outline

Compression

Lossless Compression

Lossy image compression

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Image Compression

Images and videos in raw (uncompressed) format have very large

data
1 Image/Video Size
2 One page of text 16 kb
3 Standard 640 × 480, 24-bit colour image 7.3 Mb
4 FHD 1920 × 1080, 24-bit colour image 49.3 Mb
5 Animation 50 Mbps
6 Standard video 227.6 Mbps
Transmission and storage of such large data is not easy. So
compression techniques are used.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Principles of Compression

Compression is achieved by exploiting one or more of the following

(a) Coding redundancy
(b) Spatial and Temporal redundancy
(c) Information irrelevant to human visual system

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Types of compression

(a) Lossless compression: reversible,information preserving. Text

and binary images.
(b) Lossy compression: irreversible. Most image and video
compression methods.
(c) Nearly lossless compression: for medical images.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Goals of compression

(a) Given a data amount, achieve the best quality.

(b) Given allowed distortion, minimize the data amount.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Some distortion measures

N−1
(a) Mean absolute error, MAE= N1 |x [i] − y [i]|.
P
i=0
N−1
(b) Mean squared error, MSE= N1 |x [i] − y [i]|2 .
P
i=0
!
A2
(c) Peak signal to noise ratio, PSNR=10 log10 where A
MSE
is peak intensity value.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Some compression standards names

(a) For images: CCITT group 3, CCITT group-4, JBIG1, JBIG2,

JPEG, JPEG-LS, JPEG-2000, BMP, GIF, PDF, PNG, TIFF
etc.
(b) For video: DV, H.261, H.262, H.263, H.264, MPEG-1,
MPEG-2, MPEG-4, MPEG-4 AVC, AVS, HDV, M-JPEG,
VC-1, WMV9 etc.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Average length and entropy

Consider a source of N discrete symbols Xn , n = 1, 2 . . . , N each

taking one of M discrete values.

P(Xj = an ) = pn , n = 1, 2, . . . , M

If the number of bits used to represent an are equal to l(n) then

average number of bits per symbol is
M
X
Lavg = pn l(an )
n=1

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Average length and entropy

Assuming symbols to be IID, the entropy is given by

M
X
H =− pn log2 pn
n=1

Shannon’s source coding theorem says that a code can be designed

so that
− log2 pn ≤ l(an ) ≤ − log2 pn + 1
Multiplying by pn and summing we get that

H ≤ Lavg ≤ H + 1

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Huffman code

Shannon’s theorem gives a bound but does not show how to

achieve this. Huffman coding algorithm gives a code that satisfies
this bound. It is instantaneous,uniquely decodable block code.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example Huffman code

Symbol Probability Code

a2 0.4 1
a6 0.3 00
a1 0.1 011
a4 0.1 0100
a3 0.06 01010
a5 0.04 01011
Lavg = 2.2 bits/symbol. While log2 6 = 2.585 and H = 2.14 bits.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example vector Huffman code

Symbol Probability Code

a 2/3 1
b 1/6 01
c 1/6 00
Lavg = 4/3 bits/symbol.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example vector Huffman code

Symbol Probability Code

aa 4/9 1
ab 1/9 011
ac 1/9 0101
ba 2/9 0100
bb 1/36 00111
bc 1/36 00110
ca 1/9 000
cb 1/36 00101
cc 1/36 00100
Lavg = 1.27 bits/symbol.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Disadvantages of Huffman code

(a) For large alphabet size M requires large calculations.

(b) Requires knowledge of probabilities.
(c) Coding tables also needed to be sent.
(d) JPEG, MPEG etc have default Huffman coding tables.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Some other coding schemes

(a) Arithmetic coding.

(b) Golomb and Golomb-Rice coding.
(c) Lempel, Ziv, Welch (LZW) coding.
(d) Run length coding.
(e) Predictive coding

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Run length coding

Identical consecutive symbol values are called runs. A run length

pair indicates value of new intensity and run length. Most useful
for binary images.
(a) 1-D run length coding: Assume that first intensity is always 1,
run length may be zero. Alternate code words indicate runs of
one and zeroes. A special code indicates end of line (EOL).
Run lengths are Huffman coded and runs of one and zeroes
may use different Huffman codes.
(b) 2-D run length codes: Uses relative addresses relative to
previous line. For first line previous line is assumed to be all
ones.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Bit plane coding, Symbol based coding

Bit plane coding: Images with more than two intensities can be
coded using bit plane coding. An image with M bit intensity value
can be considered as composed of M binary images by taking bits
with same weightages. Intensities are generally Gray coded. BMP
uses run length coding.

Symbol based coding: An image is represented as collection of

frequently occurring sub-images called tokens or symbols. Each
symbol is stored in a dictionary. An image is coded as sequence of
triplets, first two codes give position information and third code
gives location in dictionary.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Predictive coding

Similar to DPCM.
(a) Current pixel value can be predicted from neighboring pixels.
(b) Prediction error has nonuniform distribution near zero and can
be specified with less number of bits.
(c) Prediction can be based on current line (1-D) prediction) or
based on current and previous line (2-D prediction).
(d) In video coding motion compensated residuals are codded,
giving large compression ratio.
(e) Prediction can be done using various criteria like linear
minimum mean squared error (LMMSE) criteria or least
squares criteria.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

+ Symbol Compressed
Input +
Encoder image
-

Predictor Quantizer

Symbol + Decompressed
Input +
decoder image
+
Predictor+
quantizer

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Transform coding

Lossless image compression has generally small values. For large

compression we go for lossy compression. Block diagram of
transform based compression is shown below:

Input img Transformer Quantizer Encoder Compressed img

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

Image block-128

54 68 71 73 75 73 71 45
47 52 48 14 20 24 20 -8
20 -10 -5 -13 -14 -21 -20 -28
-13 -18 -18 -16 -23 -19 -27 -28
-24 -22 -22 -26 -24 -33 -20 -23
-29 -13 3 -24 -10 -42 -41 5
-16 26 26 -21 12 -31 -40 23
17 30 50 -5 4 12 10 5

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

2-D DCT of the block

31.00 51.70 1.16 -24.58 -12.00 -25.75 11.96 23.28

113.57 6.97 -13.90 43.20 -6.09 35.59 -13.36 -13.00
195.58 10.13 -8.66 -2.93 -28.98 -7.93 0.87 9.55
35.87 -24.30 -15.57 -20.79 11.64 -19.10 -8.53 0.51
40.75 -20.55 -13.66 17.06 -14.25 22.38 -4.89 -11.36
7.19 -13.57 -7.59 -11.94 18.25 -16.26 -1.41 -3.50
-1.45 -13.32 -0.87 1.32 10.38 16.07 4.41 1.11
-6.77 -2.83 4.11 1.11 10.55 -2.73 -3.23 1.57

Low-low coefficients are bigger than high-high coefficients.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

Normalization matrix

16 11 10 16 24 40 51 61
12 12 14 19 26 58 60 55
14 13 16 24 40 57 69 56
14 17 22 29 51 87 80 62
18 22 37 56 68 109 103 77
24 35 55 64 81 104 113 92
49 64 78 87 103 121 120 101
72 92 95 98 112 100 103 99

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

 
DCT [i,j]
QDCT [i, j] = round N[i,j]

2 5 0 -2 0 -1 0 0
9 1 -1 2 0 1 0 0
14 1 -1 0 -1 0 0 0
3 -1 -1 -1 0 0 0 0
2 -1 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

Only 19 out of 64 remain non-zero.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

Zig-zag Scanning pattern

0 1 5 6 14 15 27 28
2 4 7 13 16 26 29 42
3 8 12 17 25 30 41 43
9 11 18 24 31 40 44 53
10 19 23 32 39 45 52 54
20 22 33 38 46 51 55 60
21 34 37 47 50 56 59 61
35 36 48 49 57 58 62 63

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Example using 2-D DCT

Run-length symbol representation:

Code ={2, (0, 5), (0, 9), (0, 14), (0, 1), (1, −2), (0, −1), (0, 1),
(0, 3), (0, 2), (0, −1), (0, −1), (0, 2), (1, −1), (2, −1),
(0, −1), (4, −1), (0, −1), (0, 1), EOB}

This is Huffman coded using coding tables. Compression of the

order of 6 is obtained.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

JPEG versus JPEG-2000

Pros of JPEG
(a) Low complexity
(b) Memory efficient
(c) Reasonable coding efficiency
Cons of JPEG
(a) Single resolution
(b) Single quality
(c) No target bit rate
(d) Blocking artifacts at low bit rate
(e) No lossless capability
(f) Poor error resilience
(g) No region of interest

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

JPEG versus JPEG-2000

JPEG-2000
(a) Improved coding efficiency
(b) Full quality scalability from lossless to lossy at different bit
rates.
(c) Spatial scalabilility
(d) Improved error resilience
(e) Regions of interest can be coded with better quality
(f) More demanding in memory and computation resources

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Scalability

(a) Same image may be accessed by users with different link

speeds and display qualities. High resolution display with high
speed link or small display with mobile radio link.
(b) We can have different versions for each desirable bit rate and
quality. This is not efficient.
(c) Scalability: Asingle bit stream that can be accessed and
decoded partially.
There are two kinds of scalabilities
(a) Quality scalability: Images having different qualities. Sizes
may be same or different.
(b) Spatial scalability: Images have different sizes

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Wavelet coding

The block diagram is same as before. Instead of DCT we now use

wavelet transform. Spatial scalability is achieved by reconstructing
only the approximate image. Quality scalability is achieved by
decoding only the partial bit planes.

Input Compressed
Transformer Quantizer Encoder
image image

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Wavelet coding

Commonly used wavelets are Daubechies 9/7 bi-orthogonal wavelet

for lossy compression and LeGall 5/3 bi-orthogonal wavelet for
lossless compression. LeGall 5/3 bi-orthogonal wavelet can be
implemented using integer operations only.
Instead of square blocks standard allows for rectangular blocks
called tiles. Each tile is coded separately.

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

Wavelet coding
(a) Quantization: Each sub-band may use different step size. It
can be skipped for lossless coding.
(b) Entropy coding: Bit plane coding is used. Most significant bit
plane is coded first. Acontext based arithmetic coding is used.
Quality scalability is achieved by decoding some bit planes
starting from MSB.
(c) Region of interest (ROI): Selected region is coded with high
accuracy (e.g. face)
There are two kinds of scalabilities
(a) Quality scalability: Images having different qualities. Sizes
may be same or different.
(b) Spatial scalability: Images have different sizes

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

G. Sharma Compression
 Compression
Lossless Compression
Lossy image compression

G. Sharma Compression