Image Compression
JPEG
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 NN
1) The blocks are transformed with 2-D DCT
2) DCT coefficients are quantized
3) The quantized coefficients are encoded
JPEG: Encoding and Decoding
Source FDCT Q u an tizer E n tro p y Compressed
Image Data E n co d er Image Data
8x8 blocks
Table Table
Specifications Specifications
Compressed Entropy
D equantizer IDCT Reconstructed
Image Data Decoder Image Data
Table Table
Specifications Specifications
Divide image into NN blocks
Input image 8x8 block
2-D DCT basis functions: N=8
Low High
Low Low
8x8 block
High High
Low High
2-D Transform Coding
y00 + y23
y01 y10 y12
...
1-D DCT basis functions: N=8
u=0 u=1 u=2 u=3
1.0 1.0 1.0 1.0
0.5 0.5 0.5 0.5
0 0 0 0
-0.5 -0.5 -0.5 -0.5
-1.0 -1.0 -1.0 -1.0
u=4 u=5 u=6 u=7
1.0 1.0 1.0 1.0
0.5 0.5 0.5 0.5
0 0 0 0
-0.5 -0.5 -0.5 -0.5
-1.0 -1.0 -1.0 -1.0
1
N 1
2 j 1k N
for k 0
x j k C k cos
k
k 0 2 N 2
N
for k 1,2,..., N 1
Zig-zag ordering of DCT coefficients
DC: Direct current
AC: Alternating current
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.
Example of DCT for image block
Matlab: y=dct(x)
Distribution of DCT coefficients
DC coefficient AC coefficient
DC: uniformly distributed
AC: distribution resembles Laplacian pdf
Bit allocation for DCT coefficients
• Lossy operation to reduce bit-rate
• Vector or Scalar Quantizer?
• Set of optimal scalar quantizers?
• Set of scalar quantizers with fixed quantization tables
Bit allocation for DCT coefficients
Minimize the total distortition D
N 2 2 bi
D min hi i 2
{bi }
i 1
See Lecture 10
N
subject to b
i 1
i B
here bi is number of bits for coefficient yi,
B is a given total number of bits,
hi
1
12
p ( x) dx
i
13 3
Optimal bit allocation for DCT coefficients
Solution of the optimization task with Lagrange multiplier
method:
B 1 i 1 2
hi
Bitrate: bi log 2 2 log 2
N 2 2 H
Distortion: D NH 2 2 B N
1N 1N
N 1
2 N 1
where k
2
; H hk
k 0 k 0
Minimal distortion
Distortion: D NH 2 2 B N
1N
N 1
2
where k
2
k 0
Distortion D is minimal, if 2 is minimal.
Product of diagonal elements is greater than or equal
to the determinant of the (positive semidefinite) matrix.
Equality is attained iff the matrix is diagonal.
KLT provides minimum of 2 (and minimum of distortion D)
among other transforms!
Default quantization matrix Q
yyqq(k,l)=round[y(k,l)/Q(k,l)]
(k,l)=round[y(k,l)/Q(k,l)]
Examples: 236/16 15 Matlab: Qy=quant (y)
-22/11 -2
Quantization of DCT coefficients: Example
Ordered DCT coefficients: 15,0,-2,-1,-1,-1,0,0,-1,-1, 54{’0’}.
Dequantization
zz(k,l)=y
(k,l)=yqq(k,l)·Q(k,l)
(k,l)·Q(k,l)
Examples: 236/16 15
-22/11 -2
Matlab: z=dequant (Qy)
Original DCT block
Inverse DCT
See: x=idct(y)
Original block
Encoding of quantized DCT coefficients
• Ordered data: 15,0,-2,-1,-1,-1,0,0,-1,-1, 54{’0’}.
• Encoding:
DC: ?
AC: ?
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)
Performance of JPEG algorithm
8 bpp 0.6 bpp
0.37 bpp 0.22 bpp
Compression of color images
RGB vs YCbCr
• 24 bits RGB representation: apply DCT for each
component separately
- does not make use of the correlation between color
components
- does not make use of lowe sensitivity of the human eyes
to chrominance component
• Convert RGB into a YCbCr representation: Y is luminance,
and Yb, Yc are chrominance
- Downsample the two chrominance components
RGB YCbCr conversion
Luminance Y and two chrominances Cb and Cr
Chrominance subsampling
[Link] [Link] [Link] [Link]
1:1 2:1 Hor 4:1 Hor 2:1 Hor&Vert
Y pixel Cb and Cr pixel
Quantization of DCT coefficients
For illuminance For chrominance
Performance of JPEG algorithm
• Grayscale 8 bits images:
- 0.5 bpp: excellent quality
• Color 24 bits images:
- 0.25-0.50 bpp: moderate to good
- 0.50-0.75 bpp: good to very good
- 0.75-1.00 bpp: excellent, sufficient for most applications
- 1.00-2.00 bpp: indistiniguishable from original
JPEG JPEG2000
For illuminance
JPEG: 0.25 bpp JPEG2000: 0.25 bpp
