Mapping GOPS in an Improved DVC to H.264 Video Transcoder - Advanced Concepts for Intelligent Vision Systems - page 28

displays the average of the times reduction of the four H.264 QP points under study
compared to the reference transcoder which is composed by the full WZ decoding and
H.264 encoding algorithms.
Table 1 shows the TR of the proposed transcoder for 15fps. TR is over 71% with
negligible rate-distortion loss compared to the full complex reference transcoder.
On the other hand, Table 2 includes the results for 30fps sequences. It shows a
similar TR (over 72%) improving the RD performance with respect to 15fps sequences.
Although longer DVC GOP patterns may seem to offer a lower performance, it does
not happen so and the performance is almost the same. The approach presented here is
independent of the DVC GOP size employed. This is because of the way of our
algorithm (depicted in section 3) uses the incoming MVs. For longer DVC GOP sizes,
the previous MV generated between two K frames are ignored and the algorithm always
uses the MV generated in the last step in the WZ decoding process. The side
information in this last step is always formed by two frames with distance two between
them. These frames can be K frames or frames that have been reconstructed through the
entire WZ decoding algorithm which have been reconstructed and improved using the
parity bit information sent by the encoder. Therefore, the quality of these frames is
better than the original ones which were discarded in the first step of the algorithm.

Table 2. Performance of the proposed transcoder for 15fps sequences

RD performance of the WZ/H.264 video transcoder - 30fps
Sequence GOP TR (%)

Δ PSNR (dB) Δ Bitrate (%)

2 -0.043 1.18 74.29

Foreman 4 -0.038 0.97 74.42

8 -0.044 1.23 74.69

2 -0.004 0.11 65.84

Hall 4 -0.005 0.09 65.81

8 -0.004 0.08 65.90

2 -0.018 0.47 77.35

CoastGuard 4 -0.019 0.49 77.12

8 -0.022 0.59 76.82

2 -0.073 2.21 70.77

Soccer 4 -0.082 2.22 71.73

8 -0.074 2.08 72.05

mean -0.035 0.98 72.23

In addition, quality values were measured in SSIM terms. SSIM is an improvement
of PSNR [11] which is calculated considering luminance similarity, contrast similarity
and structural similarity. As it is shown in Figures 6 and 7, using QP = 28, 32, 36 and
40 there are no significant differences of quality and the bit rate obtained by the
H.264 reference and our proposed. Similar RD results are obtained comparing with
PSNR, as it is shown for 15 fps in Figure 8. As expected, different GOPs do not have
an important influence on the quality and the bit rate obtained by both versions.
Furthermore, for 30 fps better results are obtained due to the higher frequency.