SCT Video Compression
Paul Cockshott
Faraday Partnership

Objective
To develop a video compressor suitable for use on GSM (8Kbit) data channels
To support colour images at 12.5 frames per second.
The compressor and decompressor to run simultaneously in software on a 200mhz processor.

Customer Requirements
Compressor developed for Orange and Acorn, intially tailored to the StrongARM processor.
Target device, hand held mobile video phones.
No use of custom video compressor chips.

Technical features
Efficient use of bandwidth ( high compression ratios)
Low computational complexity compressor.
Minimal use of floating point arithmetic.
Simple decompression.

High compression ratio
Image first converted to YUV then
Vector quantized using high quality codebooks.
Blocks of image replaced by best matching codebook entries.
Motion compensation also used

Example codebooks

Multiple Scales of Blocks
Use of codebooks from 32x32 down to 4x4 using a modified quad tree with block occlusion
This contrasts with fixed sized blocks in MPEG (8x8).
Larger blocks give higer compression.

Block Occlusion
Quad tree involves total replacement
Block occlusion allows selective replacement
Half as many bits required.

CPU efficient
Vector quantization by table lookup
16 memory accesses to find best matching vector.

Code table design
Code tables constructed by recursive subdivision of probability density functions

Subdivision of Probabilty Map
Subdivision ensures that codes for like patterns have similar numerical values
Brighter patches have higher codes
This allowes recursive application of algorithm

Algorithm for PDF splitting
Determine axis of maximun variance
Split PDF through centroid perpendicular to this axis.
Assign leading 0 bit of code to lower half, 1 to upper half
Recurse on each sub portion of the PDF

Strength of SCT

Key Benefits
Videophones over GSM
Suitable for streaming video over the web decompressor in Java
Low bandwidth Digital TV (200 K baud), allows low powered satellite transmiters to send TV to wider areas.