GNR602

Advanced Methods in Satellite Image

Processing
Prof. B. Krishna Mohan
CSRE, IIT Bombay
bkmohan@csre.iitb.ac.in

Slot 13

Lecture 23 JPEG Image Compression

 IIT Bombay Slide 1
Lecture 23 Image Compression

Contents of the Lecture

Principles of Image Compression
Lossy v/s Lossless Image Compression
JPEG Image Compression
Rate Controlled JPEG Compression

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 2

What is Data and Image

Compression?
• Data compression is the representation of
information in a compact form.
• Data is a sequence of symbols taken from a discrete
alphabet.
• Still image data, that is a collection of 2-D arrays (one
for each color plane) of values representing intensity
(color) of the point in corresponding spatial location
(pixel).

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 2a
Why is an image compressible?
• Statistical redundancy:
1) Spatial correlation
a) Local - Pixels at neighboring locations have similar
intensities.
b) Global - Reoccurring patterns.
2) Spectral correlation – between color planes.
3) Temporal correlation – between consecutive frames.
• Tolerance to fidelity:
1) Perceptual redundancy.
2) Limitation of rendering hardware.
GNR602 Lecture 23 B. Krishna Mohan
Compression Ratio

compression

Compression ratio:
Relevant Data Redundancy

Example:
 Types of Data Redundancy
(1) Coding Redundancy
(2) Interpixel Redundancy
(3) Psychovisual Redundancy

• Data compression attempts to reduce one or

more of these redundancy types.
 Coding - Definitions
• Code: a list of symbols (letters, numbers, bits
etc.)
• Code word: a sequence of symbols used to
represent some information (e.g., gray
levels).
• Code word length: number of symbols in a
code word.
Coding - Definitions (cont’d)

N x M image
rk: k-th gray level
l(rk): # of bits for rk
P(rk): probability of rk

Expected value: E ( X )   xP( X  x)

x
 Coding Redundancy

• Case 1: l(rk) = constant length

Example:
 Coding Redundancy (cont’d)

• Case 2: l(rk) = variable length

variable length

Total number of bits: 2.7NM

 Interpixel redundancy
• Interpixel redundancy implies that pixel values are
correlated (i.e., a pixel value can be reasonably predicted
by its neighbors).
histograms

auto-correlation


f ( x) o g ( x )   f ( x) g ( x  a ) da


auto-correlation: f(x)=g(x)
 Interpixel redundancy (cont’d)
• To reduce interpixel redundancy, some kind of transformation
must be applied on the data (e.g., thresholding, DFT, DWT)

Example:

threshold
original

11 ……………0000……………………..11…..000…..

thresholded
(1+10) bits/pair
 Psychovisual redundancy

• The human eye is more sensitive to the lower

frequencies than to the higher frequencies in
the visual spectrum.

• Idea: discard data that is perceptually

insignificant!
Psychovisual redundancy (cont’d)

Example: quantization
256 gray levels 16 gray levels 16 gray levels + random noise

add a small pseudo-random number

C=8/4 = 2:1 to each pixel prior to quantization
 Measuring Information
• A key question in image compression is:

“What is the minimum amount of

data that is sufficient to describe
completely an image without
loss of information?”
• How do we measure the information content
of an image?
Measuring Information (cont’d)

• We assume that information generation is a

probabilistic process.
• Idea: associate information with probability!
A random event E with probability P(E)
contains:

Note: I(E)=0 when P(E)=1

 How much information
does a pixel value contain?

• Suppose that gray level values are

generated by a random process, then rk
contains:
units of information!

(assume statistically independent random events)

 How much information does an image
contain?
• Average information content of an image:
L 1
E   I ( rk ) P(rk )
k 0

using

units/pixel
Entropy:
(e.g., bits/pixel)
 Redundancy

• Redundancy:
(data vs info)

where:

Note: if Lavg= H, then R=0 (no redundancy)

 Entropy Estimation
• It is not easy to estimate H reliably!

image
 Entropy Estimation (cont’d)

• First order estimate of H:

Lavg = 8 bits/pixel R= Lavg- H

The first-order estimate provides only a

lower-bound on the compression that can be
achieved.
 Estimating Entropy (cont’d)

• Second order estimate of H:

– Use relative frequencies of pixel blocks :
image
Differences in Entropy Estimates

• Differences between higher-order estimates of

entropy and the first-order estimate indicate the
presence of interpixel redundancy!

• Need to apply some transformation to deal with

interpixel redundancy!
Differences in Entropy Estimates
(cont’d)
• Example: consider pixel differences

16
Differences in Entropy Estimates
(cont’d)
• What is the entropy of the pixel differences
image?

(better than the entropy of the original image

H=1.81)

• An even better transformation should be

possible
since the second order entropy estimate is lower:
 Image Compression Model

We will focus on the Source Encoder/Decoder only.

 Image Compression Model (cont’d)

• Mapper: transforms data to account for interpixel redundancies.

Image Compression Model (cont’d)

• Quantizer: quantizes the data to account for psychovisual

redundancies.
Image Compression Model (cont’d)

• Symbol encoder: encodes the data to account for coding

redundancies.
 Image Compression Models
(cont’d)

• The decoder applies the inverse steps.

• Note that quantization is irreversible in

general.
 Fidelity Criteria

• How close is to ?

• Criteria
– Subjective: based on human observers
– Objective: mathematically defined criteria
IIT Bombay Slide 3
How are images compressed?
• Lossless compression: reversible, information preserving
• text compression algorithms,
• binary images, palette images

• Lossy compression: irreversible

• grayscale, color, video
• Near-lossless compression:
• medical imaging, remote sensing.
• Questions to answer before compression
• 1) Why do we need lossy compression?
• 2) When do we use only lossless compression?

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 3a

Lossy vs. Lossless compression

• Achievable compression
– Lossless: Modest 1:2 to 1:3 at best
– Lossy: Can be as high as 1:100

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 4
Lossy vs. Lossless compression
100%
Example
80%

60%

40%
53.4 %
20%

6.7 %
4.3 %

0%

IMAGE: CCITT-3 LENA LENA

TYPE: binary gray-scale gray-scale
METHOD: JBIG JPEG JPEG
(lossless) (lossless) (lossy)

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 4a

Rate measures
size of the compressed file C
Bitrate: 
bits/pixel pixels in the image N

Compression ratio: size of the original file N k


size of the compressed file C

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 5

Distortion measures
N
1
Mean average error (MAE): MAE 
N
y
i 1
i  xi
N
1
Mean square error (MSE): MSE  
 i i
y  x 2

N i 1

Signal-to-noise ratio (SNR): SNR  10  log10  2 MSE  

(decibels)

Peak-signal-to-noise ratio (PSNR): PSNR  10  log10 A2 MSE  

(decibels)
A is amplitude of the signal: A = 28-1=255 for 8-bits signal.

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 5a

Lossy image compression

• DPCM: Prediction error quantization
• Block Truncation Coding (BTC)
• Vector Quantization (VQ)
• Transform Coding (DCT, JPEG)
• Subband Coding
• Wavelet Coding (JPEG2000)
Data Bitstream
Transformation Quantization Encoding

Model

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 6a

Illustration of PCT – transformation highlights low variance

bands for discarding; 1st row original, 2nd row PC 1,2,3

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 6

Transform Domain Compression

•The main goal of transformation is decorrelation of data.

X Y Q(Y) C
Transformation Quantization Encoding

Model

GNR602 Lecture 23 B. Krishna Mohan

 Lossy Compression

• Transform the image into some other domain

to reduce interpixel redundancy.
Example: Fourier Transform

Note that the

magnitude of the FT
decreases, as u, v
increase!

K << N
K-1 K-1
 Transform Selection

• T(u,v) can be computed using various

transformations, for example:
– DFT
– DCT (Discrete Cosine Transform)
– KLT (Karhunen-Loeve Transformation) or Principal
Component Analysis (PCA)
• JPEG uses DCT for handling interpixel
redundancy.
 DCT (Discrete Cosine Transform)

Forward:

Inverse:

if u=0 if v=0
if u>0 if v>0
 DCT (cont’d)
• Basis functions for a 4x4 image (i.e., cosines of
different frequencies).
 DCT (cont’d)
Using DFT WHT DCT
8 x 8 sub-images
yields 64 coefficients
per sub-image.

Reconstructed images
by truncating
50% of the
coefficients

DCT is a more
compact
transformation!

RMS error: 2.32 1.78 1.13

 DCT (cont’d)

• Sub-image size selection:

Reconstructions

original 2 x 2 sub-images 4 x 4 sub-images 8 x 8 sub-images

IIT Bombay Slide 7
2-D DCT
N 1 N 1
 2i  1k   2 j  1l 
C k , l    k  l  xi , j  cos   cos
j 0 i 0  2 N   2 N 

• The 2-D DCT is performed as two sequential 1-D DCTs

• Complexity of DCT is O(N logN) instead of O(N4)

Image y: 1-D DCT x: 1-D DCT 2-D DCT

GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 8
2-D DCT: N=4
v
0 1 2 3

GNR602 2-D Lecture 23 functions

DCT basis B. Krishna Mohan
IIT Bombay Slide 9
Zig-zag DCT coefficients ordering

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 10

Where is compression?
• DCT is reversible transformation:
Y=TX  X=T-1Y
• Where is compression?
Data decorrelation and energy compactness
 Quantization (lossy operation)
 Statistical encoding

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 11
Example

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 12
JPEG
• JPEG = Joint Photographic Experts Group

• Lossy coding of continuous tone still images (color and

grayscale)
• Based on Discrete Cosine Transform (DCT):
0) Image is divided into block NN
1) The blocks are transformed with 2-D DCT
2) DCT coefficients are quantized
3) The quantized coefficients are encoded

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 13
JPEG: Encoding and Decoding

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 14
2-D DCT basis functions: N=8
Low High
Low Low

8x8 block

High High
GNR602
Low Lecture 23 B. High
Krishna Mohan
IIT Bombay Slide 15
2-D Transform Coding

y00 + y23
y01 y10 y12

...
DC term High frequencies 
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 16
Zig-zag ordering of DCT coefficients

Converting a 2-D matrix into a 1-D array, so that the

frequency (horizontal and vertical) increases in this order
and the coefficents variance are decreasing in this order.
GNR602 Lecture 23 B. Krishna Mohan
 JPEG Steps

4. Quantize the coefficients (i.e., reduce the

amplitude of coefficients that do not
contribute a lot).

Q(u,v): quantization table

 One way to generate the
quantization table
• Quantization Table Q[i][j]
Example
IIT Bombay Slide 17

Example of DCT for image block

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 18

Default quantization matrix Q

Examples: 236/16  15
-22/11  -2
yq(k,l)=round[y(k,l)/Q(k,l)]
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 19

Quantization of DCT coefficients:

Example

Ordered DCT coefficients: 15,0,-2,-1,-1,-1,0,0,-1,-1, 54{’0’}.

GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 20
Dequantization

z (k,l)=yq(k,l)·Q(k,l)

Examples: 236/16  15
-22/11  -2

GNR602 Lecture 23 B. Krishna

Original Mohan
DCT block
IIT Bombay Slide 21
Inverse DCT

Original
GNR602 Lecture 23 B. Krishna Mohan
Original block
IIT Bombay Slide 22
Encoding of quantized DCT coefficients

• Ordered data: 15,0,-2,-1,-1,-1,0,0,-1,-1, 54{’0’}.

• Encoding:  DC: ?  AC: ?

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 23
Encoding of quantized DCT
coefficients
• DC coefficient for the current block is predicted of
that of the previous block, and error is coded using
Huffman coding
• AC coefficients:
(a) Huffman code, arithmetic code for non-zeroes
(b) run-length encoding: (number of ’0’s, non-’0’-symbol)

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 24

Suitability of JPEG Method

• The disadvantage with the present JPEG
compression technique is that the
compression rate is arbitrary.
• As the main source of compression is the
quantization table, an optimal quantization
table can help to achieve a desired
compression rate with good image quality.
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 25a

JPEG LOSSY COMPRESSION

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 25

Performance of JPEG algorithm

8 bpp 0.6 bpp

0.37 bpp 0.22 bpp

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 26

Rate control in JPEG Method

• When is a precise compression rate essential?
– Maximum quality to be preserved
– Full capacity of communication channel to be
utilized
– Data rate should not exceed the channel capacity
Key requirement in space applications

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 26a
How to control rate of
compression with JPEG?
• Quality factor is the only way to compress less
or more with JPEG.
• How can the desired rate of compression be
specified in JPEG?
• What is the implication when the rate of
compression is user specified on JPEG
technique?
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 27

Quality Factor in JPEG

• The information loss happens during the
quantization phase
• In normal case quantization table coefficients
are scaled up or down based on user-specified
• The precise rate cannot be predicted based on
the quality factor

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 33

Uncompressed CR = 3.5
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 34

Uncompressed CR =3.5

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 35

Uncompressed CR = 3.5

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 36

Uncompressed CR = 3.5

GNR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 37

Uncompressed CR = 3.5
GNR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 39

POWAI (8 BIT)
(Large Size)

NR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 40

POWAI (8 BIT); (Large Size) CR=6

NR602 Lecture 23 B. Krishna Mohan

IIT Bombay Slide 41

Original CR = 6
NR602 Lecture 23 B. Krishna Mohan
IIT Bombay Slide 42
Sample Quantization Table CR=6
2 8 9 13 13 21 22 41
8 10 12 13 21 22 40 43
10 12 15 20 23 40 44 55
11 15 20 23 39 46 54 55
17 17 25 35 46 54 56 83
17 25 35 49 51 57 63 88

28 32 50 51 59 61 104 115

30 50 50 59 61 105 112 127

NR602 Lecture 23 B. Krishna Mohan
 1 1 1 1 1 2 3 5
1 1 1 1 2 3 5 5
1 1 2 2 3 5 5 7
1 2 2 3 5 5 6 7
2 2 3 4 6 6 7 23
2 3 4 6 6 7 20 32
3 4 6 6 9 16 38 68
Uncompressed MUMBAI(11 BIT) CR = 3.5 4 6 6 12 15 65 67 71

2 8 9 13 13 21 22 41

8 10 12 13 21 22 40 43

10 12 15 20 23 40 44 55

11 15 20 23 39 46 54 55

17 17 25 35 46 54 56 83

17 25 35 49 51 57 63 88

28 32 50 51 59 61 104 115

30 50 50 59 61 105 112 127

CR = 6 CR = 9.5
 IIT Bombay Slide 43

Summary
• JPEG compressors and decompressors can be used for
compressing remotely sensed data as the Compression ratio
(CR) will seldom be high and hence JPEG gives satisfactory
performance in terms of maintaining image characteristics.
• As JPEG works with 8 by 8 block of pixels at a time the
implementation in hardware for JPEG is amenable to onboard
satellite platforms
• The JPEG process maintains a small Mean Squared Error
(MSE) at different compression ratios tested in the range
from 3.0 to 15.0.

NR602 Lecture 23 B. Krishna Mohan

Contd…