JPEG - PowerPoint PPT Presentation

JPEG

Hao Jiang

Computer Science Department

Sept. 27, 2007

• JPEG: Joint Photographic Expert Group — an international standard in 1992.
• Works for both color and grayscale images.
• Targets at natural images.
• Applications include satellite, medical imaging, general photography ...

8x8 block

DCT

Quantization

Quantization

tables

Header

DC

Entropy

coding

Tables

Difference coding

Data

Run length coding

AC

• JPEG first converts RGB to YUV or YCrCb.
• Y is the luminance component (brightness).

Y = 0.299 R + 0.587 G + 0.144 B

• U and V are color components

U = B – Y

V = R - Y

Y U V

• JPEG down-samples the “color channels” by half and partitions images into 8x8 small blocks.

U

V

Y1

Y2

Y3

Y4

Intensity Component

Color components

• DCT converts each image 8x8 block into another 8x8 block.
• The energy in DCT domain is concentrated into very few coefficients.

Image

block

DCT

DCT

coefficients

Image block DCT Coefficients

A linear transfrom F(u,v) = åx=07åy=0 7 h(u,v,x,y) f(x,y)

where h(u,v,x,y) is the linear weighting function.

DCT:

The inverse transform IDCT recovers the original data from DCT

coefficients.

F(u,v) = åx=07åy=0 7 h(u,v,x,y) f(x,y)

= åy=07g(v,y)[åx=0 7g(u,x) f(x,y)]

followed

by

Horizontal 1D DCT

Vertical 1D DCT

Given u and v, h can be viewed as an image of x and y.

u

v

DCT coefficients are projection of image block into these patterns.

• Recall that quantization can be used to collapse the input into smaller number of values.
• For DCT coefficients, we want to quantize different coefficients in different details:
• DC coefficient should have the most quantization levels.
• Quantization for AC coefficients can be coarser and coarser as the frequency increases.

Y quantization table

U, V quantization table

17 18 24 99 99 99 99 99

18 21 26 66 99 99 99 99

24 26 56 99 99 99 99 99

47 66 99 99 99 99 99 99

99 99 99 99 99 99 99 99

99 99 99 99 99 99 99 99

99 99 99 99 99 99 99 99

Quantization:

Fq(u,v) = round(F(u,v)/Q(u,v))

De-quantization:

Fr(u,v) = Fq(u,v) * Q(u,v)

quantization

dequantization

• For most current JPEG encoders, we can choose a quality factor from 1 to 100.
• The method of controlling the compression quality is by scaling the quantization table. For example,

Fq(u,v) = round(F(u,v)* (quality) / Q(u,v))

D(m,n) D(m,n+1) D(m,n+2) D(m,n+3)

We compute the difference of each two successive quantized

coefficients

D(m,n) – 0 => d(m,n)

D(m,n+1) – D(m,n) => d(m,n+1)

D(m,n+2) – D(m,n+1) => d(m,n+2)

D(m,n+3) – D(m,n+2) => d(m,n+3)

• Then, we encode each DC difference value by (size, coefficient)

• 1 -1, 1
• 2 -3, -2, 2, 3
• -7 …-4, 4, …,7
• 4
• …
• 11 -2047,…,-1024,1024,…,2047

• 7 will be coded as (3, 7)
• The first number 3 is Huffman
• coded.
• The second 7 is encoded as
• 1’s complement 111.
• (-7 will be 000)

The Size Table

Can we encode the dc difference value directly?

• AC coefficients are not differentially encoded.
• Instead, we first do run-length coding.

The run length code:

(0,11), (0,-12), (1, -12),

(0, 10), (0,16), (0, 0)

Indicates all are 0

from here

Zig-zag scanning

• For AC coefficients, we now have a bunch of symbols like
• The “value” has large number of possible values in [-1023, 1023]. Direct Huffman coding is infeasible.
• Instead, we generate symbols like

(zero-run length, value)

( zero-run length, size, value)

1’s complement

Huffman coded

Start of image

Frame

End of image

…

Tables

Header

Scan

Scan

Scan

Tables

Header

Segment

Restart

Segment

….

Block

Block

Block

• The basic mode of JPEG supports sequential coding (the order is from top to bottom and left to right).
• JPEG extensions support progressive modes:
• Spectrum selection.
• Successive approximation.
• Hierarchical mode.

• JPEG 2000 is a new standard based on wavelet transform.
• JPEG2000 does not partition an image into small blocks.
• Wavelet decomposes the whole image into different “bands” and then encodes the coefficients in different bands smartly.

LL

LH

HL

HH

The wavelet decomposition of Lena image using Haar Wavelet.