Lecture 12: JPEG
Compression: Images (JPEG)
What is JPEG?
JPEG: Joint Photographic Expert Group — an
international standard since 1992.
Works with colour and greyscale images.
Up to 24 bit colour images (unlike GIF)
Target photographic quality images (unlike GIF)
Suitable for many applications e.g . satellite, medical,
general photography...
Basic idea:
The human eye is less sensitive to higher-frequency
information.
(Also less sensitive to colour than to intensity.)
Basic JPEG Compression Pip eline
JPEG compression involves the following:
Decoding — reverse the order for encoding.
Major Coding Algorithms in JPEG
The Major Steps in JPEG Coding involve:
Colour Space Transform and subsampling(YIQ).
DCT (Discrete CosineTransform).
Quantisation.
Zigzag Scan.
DPCM on DC component.
RLE on AC Components.
Entropy Coding — Huffman or Arithmetic.
We have met most of the algorithms already:
JPEG exploits them in the compression pipeline to
achieve maximal overall compression.
Quantization
Why do we need to quantise:
To throw out bits from DCT.
Example: (101101)2 = 45 (6 bits).
Truncate to 4 bits: (1011)2 = 11.
Truncate to 3 bits: (101)2 = 5.
Quantization error is the main source of Lossy
Compression.
DCT itself is not Lossy.
How we throw away bits in Quantization Step is
Lossy.
Quantization
Uniform quantization
Divide by constant N and round result
(N = 4 or 8 in examples on previous page).
Non powers-of-two gives fine control
(e.g., N = 6 loses 2.5bits)
Quantization Tables
In JPEG, each F[u,v] is divided by a constant q(u,v).
Table of q(u,v) is called quantization table.
Eye is most sensitive to low frequencies (upper left
corner), less sensitive to high frequencies (lower right
corner)
JPEG Standard defines 2 default quantization tables,
one for luminance (below), one for chrominance. E.g.:
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
Quantization Tables
Q: How would changing the numbers affect the picture?
E.g . if we doubled them all?
Quality factor in most implementations is the scaling
factor for default quantization tables.
Custom quantization tables can be put in image/scan
header.
JPEG Quantisation Example
JPEG Quantization Example (Java Applet)
Zigzag Scan
What is the purpose of the Zig-zag Scan:
To group low frequency coefficients in top of vector.
Maps 8 x 8 to a 1 x 64 vector
Differential on Pulse Code Modulation
(DPCM) DC Component
DPCM is then employed on the DC component.
Why is this strategy adopted:
DC component is large and varies, but often close to
previous value.
Encode the difference from previous 8x8 blocks—
DPCM
Run Length Encode (RLE) on AC Components
Yet another simple compression technique is applied to
the AC component:
1x63 vector (AC) has lots of zeros init
Encode as (skip, value) pairs, where skip is the number
of zeros and value is the next non-zero component.
Send (0, 0) as end-of-block sentinel value.
Huffman (Entropy) Coding
DC and AC components finally need to be represented by a
smaller number of bits (Arithmetic coding also supported in
place of Huffman coding):
(Variant of ) Huffman coding: Each DPCM-coded DC
coefficient is represented by a pair of symbols:
(Size, Amplitude)
where Size indicates number of bits needed to represent
coefficient and Amplitude contains actual bits.
Size only Huffman coded in JPEG:
Size does not change too much, generally smaller
Size occurs frequently (= low entropy so is suitable for
entropy coding),
Amplitude can change widely so coding no real
benefit.
Huffman (Entropy) Coding
Example Size category for possible Amplitude s:
---------------------------------------------------------------------------
Size Typical Huffman Code for Size Amplitude
0 00 0
1 010 -1,1
2 011 -3,-2,2,3
3 100 -7..-4,4..7
4 101 -15..-8,8..15
… … …
… … …
---------------------------------------------------------------------------
Use ones complement scheme for negative values: i.e
10 is binary for 2 and 01 for -2 (bitwise inverse). Similarly,
00 for -3 and 11 for 3.
Huffman Coding DC Example
Example: if DC values are 150, -6,5, 3, -8
Then 8, 3, 3, 2 and 4 bits are needed respectively.
Send off Sizes as Huffman symbol, followed by actual
values in bits:
(8huff, 10010110), (3huff, 001), (3huff, 101), (2huff, 11), (4huff, 0111)
where 8huff, ... are the Huffman codes for respective
numbers.
Huffman Tables can be custom (sent in header) or
default.
Huffman Coding on AC Component
AC coefficient are run-length encoded (RLE)
RLE pairs(Runlength, Value) are Huffman coded as
with DC only on Value.
So we get a triple: (Runlength, Size, Amplitude)
However, Runlength, Size allocated 4-bits each and put
into a single byte with is then Huffman coded.
Again, Amplitude is not coded.
So only two symbols transmitted per RLE coefficient:
(RLESIZEhuff byte, Amplitude)
Example JPEG Compression
Another Enumerated Example
JPEG Example MATLAB Code
The JPEG algorithm may be summarized as follows:
im2jpeg.m (Encoder) jpeg2im.m (Decoder)
mat2huff.m (Huffman coder)
m = [16 11 10 16 24 40 51 61 %JPEG normalizing array
12 12 14 19 26 58 60 55 % and zig-zag reordering
14 13 16 24 40 57 69 56 %pattern.
141722 29 5187806218 223756 68109103 772435556481 104113
924964 78 8710312112010172 929598112100103 99] *quality;
Order= [19 23101725 18114512 192633 … 4134 2720 136714
212835 42 495750 … 4336 2922 158 16 2330 3744 5158 5952 …
453831 24 3239465360 615447 404855 … 6263 5664];
[xm, xn] = size(x); %Get input size.
x = double(x) - 128; %Level shift input
t = dctmtx(8); %Compute 8x 8 DCTmatrix
%Compute DCTs of 8x8 blocks and quantize the coefficients.
y = blkproc(x,[88], ’P1*x* P2’,t, t’);y=blkproc(y, [88],
’round(x ./ P1)’, m);
JPEG Example MATLAB Code
y = im2col(y, [8 8], ’distinct’); %Break 8x8 blocks into columns
xb = size(y, 2); % Get number of blocks
y = y(order, :); %Reorder column elements
eob = max(y(:)) + 1; % Create end-of-block symbol
r= zeros(numel(y) +size(y, 2), 1); count =0;
for j = 1:xb %Process 1block (col) at a time
i = max(find(y(:, j))); % Find last non-zero element
if isempty(i) %No non zero block values
i= 0; end;
p = count + 1;q= p + i;
r(p:q) = [y(1:i, j); eob]; %Truncate trailing 0’s, add EOB,
count = count + i + 1; %and add to output vector
end
r((count + 1):end) = []; %Delete unusued portion of r
y= struct; y.size = uint16([xm xn]); y.numblocks = uint16(xb);
y.quality = uint16(quality *100); y.huffman =mat2huff(r);
Artefacts
This image is compressed increasingly more from left
to right.
Note ringing artefacts and blocking artefacts.
artefacts.
Gibb’s Phenomenon
Artefacts around sharp boundaries are due to Gibb’s
phenomenon.
phenomenon
Basically: inability of a finite combination of cosines to
describe jump discontinuities.
Gibb’s Phenomenon
Gibb’s Phenomenon
Further Information
Further standards:
Lossless JPEG: Predictive approach for lossless
compression, not widely used.
JPEG 2000: ISO/IEC 15444
Based on wavelet transform, instead of DCT, no 8×8
blocks, less artefacts.
Often better compression ratio, compared with JPEG.
Further Information
References:
http://www.jp eg.org
Online JPEG Tutorial
The JPEG Still Picture Compression Standard
The JPEG 2000 Still Image Compression Standard
