1. Lossless JPEG transcoding
Daniel Sanchez
ECE533 Fall 2006
Final Project Presentation
University of Wisconsin-Madison

2. Overview The problem of JPEG transcoding
Full-featured MATLAB JPEG codec
Lossless transcoder
Lossless rotation

3. The problem of JPEG transcoding
Basic JPEG decoder/encoder block diagrams:
Reencoding introduces errors because of the color model conversions (YCbCr<->RGB)
JPEG Decoder
DC Huffman
IDPCM DC
Level shifting
YCbCr->RGB
JPEG
File IQ
IDCT
RGB Image
AC Huffman
DeZigZag AC
JPEG Encoder DC
DPCM
DC Huffman
RGB->YCbCr
Level shifting
RGB Image
JPEG
File
DCT Q AC
ZigZag
AC Huffman

4. Implementing a JPEG codec in MATLAB
Extends the MATLAB JPEG Toolbox by Phil Sallee
Includes functions that cover the more tedious parts of the codec
Works with most images!
Full color, arbitrary size,…
Some restrictions in subsampling and color models
Implemented
JPEG Encoder
Provided by toolbox DC
DPCM
DC Huffman
RGB Image
RGB->YCbCr
Level shifting
JPEG
File
DCT Q AC
ZigZag
AC Huffman

5. Example of lossy reencoding
Original image
Error after reencoding (x15)
SNR=25.1dB

6. Lossless JPEG transcoder [1]
How to modify decoder/encoder?
Lossy JPEG Transcoder
DC Huffman
IDPCM DC
Level shifting
YCbCr->RGB
JPEG
File IQ
IDCT
RGB Image
AC Huffman
DeZigZag AC DC
DPCM
DC Huffman
RGB->YCbCr
Level shifting
RGB Image
JPEG
File
DCT Q AC
ZigZag
AC Huffman
Reencoded coefficients are
an estimation of the original ones
-> Correct that estimation!

7. Lossless JPEG transcoder [2]
JPEG Lossless Transcoder - Decoder
DC Huffman
IDPCM DC
JPEG
File
Level shifting
YCbCr->RGB IQ
IDCT
RGB Image AC
AC Huffman
DeZigZag -
RGB->YCbCr
Level shifting
DCT Q
Lossless Compression
Difference coefficients +
Lossless Decompression
Difference coefficients
DPCM
DC Huffman
RGB->YCbCr
Level shifting
JPEG
File
RGB Image
DCT Q
ZigZag
AC Huffman
JPEG Lossless Transcoder - Encoder

8. Results of lossless transcoding
Difference coefficients (contrast ↑↑)
Original image
Transcoded image
Lossy reencoding: SNR = 33.2 dB
Lossless transcoding: SNR = ∞

9. Lossless rotation Different approach: Operate on blocks directly
No need to requantize -> Lossless
Faster than decoding/encoding
To rotate an image (90º counterclockwise):
Rotate whole blocks (without modifying content)
Perform the transformation G(u,v)=F(v,u)(-1)v on each block
This rotates the contents in the spatial domain!
Transpose quantization matrices (they are not symmetric in general!) 1 2 3 4 5 6 7 8 9 3 6 9 2 5 8 1 4 7

10. Results of lossless rotation [1]
Original image

11. Results of lossless rotation [2]
Those strange manipulations actually work!
Losslessly rotated image

12. Results of lossless rotation [3]
Error when lossy rotation is applied 4 times
This time SNR=18.1 dB
Lossy rotation degrades the quality much more than reencoding!

13. Applications of lossless transcoding
Image editing (allows editors to work with lossy compressed images without degrading quality)
Image rotation and other basic transformations (cropping, flipping,…)
Image transmission/storage

14. Thanks
for your attention
Any questions?