Section 5.2 Proposed Momentum-directed Genetic Pattern Searches
5.2.1 Performance of Momentum-Directed Genetic Pattern Searches
To test the proposed algorithm, ten sequences (denoted as ‘1X’) with different MV variances are tested under the parameter settings given in Table 3-1. Moreover, to test the extreme cases, we generate ten new test sequences by skipping the even frames of these sequences, and these new sequences are denoted as ‘2X’. They are roughly the two times fast forward playback of the originals. These 20 test sequences are coded by an MPEG-4 SP@L3 encoder. The other simulation settings are the same as described in Section 3.1.
In selecting the simulation platform, our focus is whether it provides a fair and direct comparison among different ME algorithms. The H.264 scheme is a newer and very sophisticated platform. It contains many tools that affect the choice of motion vectors, such as multiple (block) mode decision and rate distortion optimization. For example, at different bit rates, the same mode
92
-decision tool can select different motion vectors. Thus, the PSNR impact due to the use of different ME algorithms may become hidden or be blurred. Thus, we adopt a simpler MPEG-4 platform on which the impact of different motion estimation algorithms can be observed more clearly.
The average number of search points (ASP) and the peak signal to noise ratio (PSNR) for various sequences and search algorithms are listed in Table 5-2 and Table 5-3, respectively. The predicted MV (PMV, defined by (2.2)) is used as the search starting point in all cases. FS denotes the full search, ERPS is proposed by [30] but we replace the MV predictors in [30] by the PMV, and PHS is proposed by [33].
The pair-wise performance comparisons in ASP and PSNR between MD-GRPS and some selected popular algorithms are given in Fig. 5-15 and Fig. 5-16. The pair-wise performance comparisons in ASP and PSNR between MD-GPHS and some selected popular algorithms are given in Fig. 5-17 and Fig. 5-18. In Fig. 5-15 and Fig. 5-17, the computing gain (CG) is defined as the ASP ratio between the original and the chosen algorithm minus one. In Fig. 5-16 and Fig.
5-18, the quality gain (QG) is the PSNR difference. The CG of MD-GRPS and MD-GPHS substantially outperforms the other popular algorithms, while their average QG is near 0.
MD-GRPS can be up to 18% faster than GRPS for very fast sequences (2X FB1024), and their PSNR values are about the same. On the average, comparing their ASP values, MD-GRPS is 7% faster than GRPS, 35% faster than ERPS, 1.39 times faster than DS, 1.76 times faster than FSS and 143 times faster than FS. And the PSNR of MD-GRPS is about the same as that of all the other search algorithms (+0.06dB ~ -0.06dB).
Similarly, MD-GPHS can be up to 13% faster than GPHS for very fast sequences (2X FB1024) and its PSNR quality is roughly at the same level. On the average, MD-GPHS is 5%
faster than GPHS, 12% faster than PHS, 69% faster than DS, 96% faster than FSS, and 101 times faster than FS. And the PSNR of MD-GPHS is about the same as those of GPHS and the
93
-non-genetic version (PHS) (+0.02dB ~ -0.05dB). When being compared to the conventional pattern search algorithms, all these three algorithms (MD-GPHS, GPHS, and PHS) have slightly PSNR drop (-0.12dB ~ -0.17dB).
Generally MD-GRPS is significantly better in speed than MD-GPHS for most test sequences.
However, MD-GPHS outperforms MD-GRPS by 3% for very fast sequences (2X FB1024). As expected in comparing Fig. 5-8 with Fig. 5-11, RWFMD-GPHS has smaller values than RWFMD-GRPS
near the outer border of the search area. In short, one algorithm beats the other in certain scenarios but none is the best for all cases. Thus, a good adaptive pattern scheme that dynamically selects the most appropriate pattern search algorithms further reduces the computational complexity.
The computation overheads of MD-GRPS and MD-GPHS are negligible. For all possible pairs of the last and the second-to-the-last successful mutation directions, we generate the corresponding search priority tables in advance. In execution of a momentum-directed algorithm, we record the last and the second-to-the-last successful mutation directions, use this direction pair to choose the search priority table and decide the search priority accordingly. A few memory access and comparisons can do all the works.
Note that, in Table 5-2, the PSNR of both ‘1X’ and ‘2X’ HL40 acquired by FS are lower than those acquired by other algorithms. Because HL40 has slight noise textures, the motion vector field produced by FS is much noiser (with larger magnitude) than those produced by the other algorithms. This phenomenon influences the matching error little but the size of motion vectors a lot, therefore, has a more significant influence on the coded picture quality, particularly for low bitrate sequences.
Table 5-2 ASP (Average Number of Search Points).
Type Sequence MD-GRPS GRPS ERPS MD-GPHS GPHS PHS DS FSS FS CT256 5.28 5.36 5.75 9.19 9.37 9.52 13.81 17.53 1024 1X
CT40 5.85 5.98 7.04 9.51 9.88 10.31 15.03 18.38 1024
94
-HL40 6.23 6.35 7.33 9.60 9.68 10.10 15.38 18.72 1024 MD96 5.94 5.98 6.83 9.58 9.65 10.02 14.85 18.37 1024 CG112 5.75 6.08 7.63 9.29 9.76 10.25 15.09 18.25 1024 FM512 6.80 7.13 8.65 9.80 10.00 10.57 16.17 19.03 1024 FM1024 6.64 6.94 8.32 9.67 9.85 10.35 15.76 18.71 1024 FB1024 10.35 11.89 16.36 11.36 12.75 14.18 22.36 22.70 1024 FG768 6.06 6.38 7.57 9.72 9.95 10.34 15.30 18.73 1024 ST1024 7.24 7.65 9.95 9.90 10.56 11.40 16.96 19.47 1024 CT256 5.43 5.62 6.35 9.26 9.51 9.74 14.15 17.72 1024 CT40 6.40 6.60 8.15 9.82 10.34 10.89 16.05 19.11 1024 HL40 6.37 6.51 7.57 9.66 9.74 10.22 15.62 18.88 1024 MD96 6.29 6.40 7.56 9.77 9.85 10.38 15.44 18.76 1024 CG112 6.73 7.36 9.54 9.74 10.64 11.48 17.04 19.57 1024 FM512 8.25 9.07 11.70 10.52 11.01 12.02 18.72 20.67 1024 FM1024 7.98 8.85 11.36 10.35 10.79 11.75 18.26 20.28 1024 FB1024 13.27 15.75 22.32 12.94 14.62 17.15 27.39 26.22 1024 FG768 6.55 7.01 8.69 9.88 10.35 10.83 16.30 19.29 1024 2X
ST1024 8.61 9.28 12.45 10.72 11.73 13.00 19.49 21.26 1024 Average 7.10 7.61 9.56 10.01 10.50 11.23 16.96 19.58 1024.00
Table 5-3 PSNR (Peak Signal to Noise Ratio).
Type Sequence MD-GRPS GRPS ERPS MD-GPHS GPHS PHS DS FSS FS CT256 39.48 39.49 39.50 39.47 39.43 39.44 39.51 39.49 39.56 CT40 31.99 32.21 32.08 31.28 31.24 31.47 31.92 31.69 32.04 HL40 34.41 34.49 34.60 34.15 34.14 34.22 34.25 34.17 33.55 MD96 40.05 40.08 40.09 39.78 39.79 39.85 39.99 39.93 39.80 CG112 29.13 29.14 29.16 29.06 29.03 29.06 29.14 29.13 29.08 FM512 34.04 34.05 34.10 33.86 33.89 33.92 34.06 34.02 34.06 FM1024 36.55 36.52 36.61 36.49 36.46 36.44 36.59 36.48 36.56 FB1024 34.92 34.87 34.88 34.85 34.73 34.87 34.93 34.94 35.28 FG768 26.18 26.17 26.19 26.14 26.15 26.17 26.18 26.16 26.20 1X
ST1024 29.16 29.39 29.31 29.31 29.42 29.33 29.44 29.35 29.48 CT256 38.63 38.65 38.68 38.52 38.52 38.51 38.60 38.72 38.95 CT40 30.15 30.28 30.22 29.30 29.22 29.54 29.94 29.73 29.81 HL40 33.25 33.31 33.38 32.91 32.95 33.02 33.07 32.93 32.33 MD96 38.66 38.66 38.66 38.39 38.37 38.44 38.60 38.57 38.41 CG112 27.43 27.43 27.53 27.33 27.23 27.34 27.50 27.45 27.37 FM512 32.36 32.34 32.45 32.16 32.19 32.23 32.38 32.35 32.42 FM1024 35.23 35.25 35.29 35.14 35.12 35.21 35.24 35.17 35.28 2X
FB1024 33.26 33.22 33.24 33.21 33.12 33.22 33.28 33.30 33.44
95
-FG768 25.52 25.51 25.53 25.46 25.42 25.48 25.53 25.49 25.51 ST1024 27.86 27.99 27.93 27.87 27.87 27.88 27.97 27.93 28.10 Average 32.91 32.95 32.97 32.73 32.71 32.78 32.91 32.85 32.86
Fig. 5-15 Performance comparisons in ASP between MD-GRPS and some popular algorithms.
Performance comparison in ASP between MD-GRPS and some popular algorithmgs
0.000
CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 Average
1X 2X
Fig. 5-16 Performance comparisons in PSNR between MD-GRPS and some popular algorithms.
Performance comparison in PSNR between MD-GRPS and some popular algorithms
-1.00
CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 Average
1X Test Sequences 2X
QG (Quality Gain)
Fig. 5-17 Performance comparisons in ASP between MD-GPHS and some popular algorithms.
Performance comparison in ASP between MD-GPHS and some popular algorithms
0.000
CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 Average
1X 2X
Fig. 5-18 Performance comparisons in PSNR between MD-GPHS and some popular algorithms.
Performance comparison in PSNR between MD-GPHS and some popular algorithms
-1.00
CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 CT256 CT40 HL40 MD96 CG112 FM512 FM1024 FB1024 FG768 ST1024 Average
1X Test Sequences 2X
QG (Quality Gain)
96