Algorithms of DVC - Distributed Multiple Description Coding - page 128

Fig. 5.8 Visual quality. ( a ) Original, ( b ) MC interpolated, ( c ) reconstructed (40 kbps), ( d )refined
reconstruct(40 kbps)

5.3 Residual DVC Based on LQR Hash

5.3.1 Motivation

Generally in DVC system, for the Wyner-Ziv frames, the quantized pixels or
transform coefficient X is channel encoded, and only the parity bits P are
transmitted. At the decoder, the error-correcting decoding using the parity bits P
and SI Y is implemented to recover X . The reconstruction is to recover the original
pixels or coefficients with the help of SI generated by the decoded key frames.
As the estimation to the main source, SI Y is very important in DSC and
DVC. Generally, the more dependent Y is, the less parity bits and the higher
compression as well as the better reconstruction can be achieved. Among some
methods, hash-based SI is the most attractive method [ 13 ].Themainideaisthat
except the Wyner-Ziv bits, the encoder also sends some representing information
on encoding to assist the motion estimation. Because the hash bits take more
information about current frame, the hash-based motion estimation is more accurate
than that without hash. On the other hand, the hash-based DVC is enhanced at
the cost of hash bits sent. Therefore, the hash should be compressed efficiently.
Specially, in [ 13 ], the compression of hash bits is based on the entropy code, so
the hash bit rate is R.h/ H.h/. Martinian et al. [ 17 ] propose use of the LQR
frame to generate efficient hash. The LQR hash refers to the low-quality version
of the current Wyner-Ziv frame. LQR hash is coded by the prediction coding
similar to the DPCM/DCT with 0-motion-searching and coarse quantization. Just
because the temporal dependence is exploited, the bit rate of LQR hash approaches,
ideally, the conditional entropy R.h/ H.h j Y/, and better rate-distortion can be
obtained. But, the LQR-based method in [ 17 ] has an apparent disadvantage. For
example, LQR hash bits and Wyner-Ziv bits are two different streams compressed
by 0-motion H.264 and Wyner-Ziv codec, respectively, for the same frame. So there
is redundancy between LQR and Wyner-Ziv bits especially at high bit rate.
On the other hand, it is shown that residual DVC can give better performance
than pure pixel DVC [ 18 ]. Residual DVC means Wyner-Ziv encoding the residual
D D W W re ,whereW re is a simple estimation to W and should be accessible both
at the encoder and the decoder. D y D Y W re is used as SI Wyner-Ziv decoding.