Reversed Complexity Video Compression

Reversed Complexity Video Compression
Jacco Taal, Inald Lagendijk, Paul Meyer
Video Encoding and Decoding are Asymmetric in
Complexity
New Video Compression Paradigm based on
Distributed Source Coding
Since the introduction of motion-compensated DCT coders
(MPEG), the belief has been that
• Video encoders are inherently complex
- Motion estimation
- Quality and rate control (quantization control)
• Decoders are “slaved” to the encoders
• Compression quality is controlled by encoder
Old situation
Even though progress has been made in reducing encoder
complexity …
- efficient motion estimation
- complexity scalable coding
scenarios exist in which encoder complexity is prohibitive.
Distributed Source Coding / Source Coding with Side
Information
Decoded video
Input video
Complex
video encoder
Simple
video decoder
New situation: Complex decoder
Decoded video
Input video
Simple
video encoder
Complex
video decoder
New situation: Complex transcoder
Decoded video
Encoder can be made simpler by just ignoring available
(side) information. This makes the processing at the
encoder simpler, while the overall performance can be close
to the case where the (side) information ís used by the
encoder.
Input video
Simple
video encoder
Complex
transcoder
Simple
decoder
Key insight (Slepian-Wolf)
H(X|Y)
Encoder
Decoder
^
X
Ideally, the bit rate - SNR performance of the three systems
should be close (e.g. close to MPEG-2 or H.264)
X
Y
Y
Wyner-Ziv-based Video Encoder
H(X|Y)
Encoder
Decoder
I
X
^
X
Me/Mc
I
X
Y
Coset definition
Syndrome encoder
Assuming sufficiently correlated Y
Information theory: X can be theoretically compressed at
a rate equal to that when the encoder too has access to Y
WZ
(Slepian-Wolf 1973)
Y
WZ Decoder
Xˆ
Coset-bits (Syndromes or
“Parities” of channel codes)
Case where Xˆ is a degraded version of X (lossy
compression) was addressed by Wyner-Ziv (1976) in a ratedistortion framework
What Makes the System “Difficult”: Research
1.
Optimal combination of signal processing transforms
and syndrome encoding
2.
Side information Y created at decoder is
• inherently less correlated to X than at encoder
side
• less constrained: no need to transmit parameters
used to generate side-information (e.g. motion
vectors)
3.
Encoder is slaved to decoder via correlation, not vice
versa as in current compression systems
All possible codes
coset #1
coset #2
coset #N
Xˆ
Y