Realtime Videocompression based on 3DDCT

1
Real-time Videocompressionbased on 3D-DCT
2
Presentation Overview

• Algorithm
• Demo
• Design
• Functional Summary

3
Basics

• decorrelation of subsequent frames(transform to
  frequency domain)
• lossy compression

4
3D-DCT transformation
5
3D-DCT transformation

• Different cube depths (1, 4, 8)
• better compression ratio
• latency
• cross fading over scene cuts

6
Decorrelation of adjacent cubes

• DC - Prediction
• differential encoding of DC coefficient of
  adjacent cubes
• especially efficient with strong quantization
• no loss in quality

• Differential - Prediction
• uses a frame as a predictor for subsequent
  frames
• improves compression ratio w/o increasing
  latency
• involves loss in quality

7
Entropy reduction / Streaming

• Quantization
• goal remove information that is less important
  to the human visual system
• suppress high frequencies
• use YUV representation

• Streaming
• goal optimizing the stream for the subsequent
  source coding
• read the cube in diagonal planes
• rearrange Y-/U-/V-channels

8
Source Coding

• Actual Compression
• goal exploit special ordering and
  probability distribution of the data stream
• Zero Run Length coding
• Variable Length coding

9
Source Coding (VLC)

• Symmetry ? code only absolute value
• Limited range ? use prefix coding
• Terminator code for word alignment

10
Image Quality

• Factors
• desired compression ratio (quantization)
• type of sequence
• static
• morphologic
• translatoric

• Measurement
• mathematical (PSNR)
• subjective (human estimation)

11
Image Quality

• Typical example for video conferencing
• static background
• morphologic motion

12
MPEG
3D-DCT
13
Image Quality

• Typical example for a movie scene
• motion blur
• fast, translatoric motion
• scene cuts

14
3D-DCT
MPEG
15
Design
16
Frontend

• Video Frontend frame buffering

17
Frontend

• Video modes
• Interlaced
• Up-/down-sampling
• Video preprocessing
• YUV conversion
• Color offset correction
• Differential Prediction

18
3D-DCT

• 3D-DCT cube extraction

19
3D-DCT

• Requirements
• DCT/IDCT, Cubedepths 1,4,8
• 72000 cubes/s (PAL)
• 13800000 1D-DCT/s

• Architecture
• Parallel processing of the three color channels
• 3D-DCT ? 364 1D-DCT ? Butterfly
• controlled via ?Code-ROM

20
Post Processing

• Stream unit continuos data flow

21
Stream processing

• Streamer
• Reordering of the cube in diagonal planes
• Separation of the Y-channel from the
  U/V-channels
• DC-Prediction

• Quantization
• Loadable 16 bit quantization table for each
  channel

• ZRL-En/Decoder
• Symmetrical run length representation

22
VLC

• Requirements
• limited range from -256 to 256 (sign symmetry)
• maximum code length 22 bit
• throughput 50 Mcodes/s

• Encode with direct lookup
• configurable memory table 256 ? 22 bit
• three onchip ROMs 3 ? 0.057 mm2

23
VLC

• Decode with direct lookup
• configurable memory 22 bit address decoder (8
  MB)

• three onchip ROMs 3 ? 0.084 mm2

A
0101011100100
101011100100
C
011100100
A
24
Functional Summary

• Single chip Video-compression system

25
Outlook

• Exploration of different quantizations (with
  respect to the human visual system)

• Improvements with adaptive features for
• quantization
• cube depth
• differential encoding

• Exploration of QoS aspects
• transmission over lossy channels
• adaptive compression ratio

26

• Physical Data
• System clock 100 MHz
• Video Interface 15 MHz
• 70000 standard cells
• 92 kbit memory
• die area 22.6 mm2
• process 0.35, 5-metal

Questions ??