Algorithms of MD - Distributed Multiple Description Coding - page 87

value is chosen from 65 to 210 for different bit rates tested, where all the coarse
information can be embedded into the fine information. By adjusting the finer
quantization value q 1 from 8 to 65, the bit rate per description varies, and q 2 is

selected as 2 for a small modification due to embedding.
For a fair comparison, we tested our proposed embedding-based coding scheme
against the following two conventional methods without such embedding, with the
same coarse and fine coders for the two parts. That is to say, the fine information and
the coarse information used for the three testing schemes are exactly the same.

1. Coarse Information Skipping : In the coarse information skipping approach, each
description only contains the fine information from the fine coder and skips the
coarse information, resulting in the same coding rate per description as in our
coarse information embedding scheme. In this way, the redundancy between the
two descriptions is minimized to favor the central decoding performance. When
only one description is received, an interpolation method used in [ 36 ] is exploited
to recover the missing part.
2. Coarse Information Appending : In the coarse information appending scheme,
the coarse information is appended after the fine information. Each description
is produced by concatenating the fine information and the coarse information
for the two parts alternatively. Note that the coarse information and the fine
information used in this scheme are the same as those in the proposed scheme.

Figures 4.31 and 4.32 plot the rate-distortion performance of side/central coding
to compare our proposed method against the other two relevant schemes on
two testing images, respectively. Both spatial and wavelet domain partitions are
considered in the comparison, where 8

8 blocks are used for the spatial splitting,

and 32 32 tree-structured blocks are for the wavelet domain splitting. In the

experiment, the bit rate of coarse information ranges from 0.05 to 0.12 bpp. As
mentioned before, in this letter, we consider all the coarse information to be
embedded into the fine information. We take the spatial domain partition as an
example for the following discussions. It can be seen from Figs. 4.31 aand 4.32 a
that, for the side coding, the proposed method outperforms the coarse information
skipping approach substantially as the bit rate increases, for example, over 3 dB gain
for “Lena” and 4 dB gain for “Barbara.” Compared with the coarse information
appending method, the proposed scheme can achieve the similar decoded quality
(PSNR) with a lower bit rate up to 12.9% rate reduction for “Lena” and 12.6% rate
reduction for “Barbara.” For the central coding, the coarse information skipping
method achieves the best central decoded results as expected, at the expense of the
poorest side decoded results described above. Figures 4.31 band 4.32 b demonstrate
that the proposed one achieves similar results (with a PSNR degradation less than
0.6% and 0.8% for “Lena” and “Barbara,” respectively) as the coarse information
skipping method in central coding, and significantly better results than the coarse
information appending scheme due to the bit rate saving. From the results, we can
clearly see that our proposed scheme achieves overall better coding results than the
two conventional methods in terms of side and central distortion-rate performance.