5. Experiment and Analysis
5.2. T Result of Proposed Bit Allocation Scheme with Weighting Threshold
5.2.2. T The result of proposed scheme with linear formula
XTable 5.8X and XTable 5.9X show the results of testing the algorithm on three sequences using Method I and Method II, respectively. The test conditions and the meaning of the fields in the table are the same as those in section 5.1 (in particular Table 5.2).
Sequence Stefan Football Bus
Type Original Modified Original Modified Original Modified PSNR 29.52022 29.52101 33.48467 33.49391 31.14866 31.14505 SSIM 0.91944 0.91938 0.87353 0.87397 0.89351 0.89323
QP 14 13~15 10 9~11 10 9~11
Total Bitrate 828 837 1177 1187 1279 1287 Luma Bitrate 655 655 841 841 1103 1103 Header Bitrate 133 141 245 256 150 157
Chroma Bitrate 39 39 89 89 25 25
Table 5.8 Average result of Method I
Sequence Stefan Football Bus
Type Original Modified Original Modified Original Modified PSNR 29.52022 29.50551 33.48467 33.46981 31.14866 31.1424 SSIM 0.91944 0.9191 0.87353 0.87381 0.89351 0.89236
QP 14 13~15 10 9~11 10 9~11
Total Bitrate 828 852 1177 1194 1279 1291 Luma Bitrate 655 655 841 840 1103 1103 Header Bitrate 133 156 245 263 150 161 Chroma Bitrate 39 39 89 89 25 25
Table 5.9 Average result of Method II
Method I perform slightly better on distinguishing between SR and USR than Method II.
Moreover, on average, Method I is slightly better than Method II on video quality too.
XTable 5.8X shows the result of a single frame by Method I. For each sequence we list two kinds of typical frames to analyze the result of visual quality and bits our proposed scheme causes. XFigure 5.4X, XFigure 5.5X and XFigure 5.6X show the result of visual quality in three sequences respectively. Symbols in these figures are as figures in section 5.1 defined.
(a) (b) Figure 5.4 Comparison of visual quality in Stefan Sequence
(a)The 43PrdP frame in Stefan. (b) The 80PthP frame in Stefan.
43PrdP frame 80PthP frame
PSNR Better Worse Better Worse
Audience 99 33 87 45
Words 28 6 16 1
Whole body 28 8 18 6
Table 5.10 Ratio of number of regions with better PSNR and worse PSNR in Stefan sequence
For the Stefan sequence, human observers may pay special attention to tennis player and the area with obvious edges such as words on the wall. On the other hand, the regions that audiences on the grandstand and the flat regions are mostly human observers are not sensitive to relatively. In XFigure 5.4X(a), the major movement in the 43PrdP frame is the tennis player moving towards right hand side. In our proposed scheme, the regions with clear and obvious directional edges will be considered structured region. As a result, performances of this kind of regions such as tennis player’s legs, words on the wall and lines on the ground are mostly enhanced. Nevertheless, the regions of audiences on the grandstand and the flat regions will
be seemed to unstructured regions since their directions of edges are complicated. So bits of these regions are usually decreased and it may cause worse visual quality.
In XFigure 5.4X(b), the 80PthP frame, the major movement in the 80PthP frame is the tennis player waving his rocket. Therefore, human eyes may notice the area of tennis player’s whole body and the area with obvious edges such as words on the wall and lines on the ground.
Performances of these kinds of regions are mostly enhanced. And similar as in XFigure 5.4X(a), bits of the regions of audiences on the grandstand and the flat regions are usually decreased and it may cause worse visual quality. XTable 5.10X shows ratio of number of regions with better PSNR and worse PSNR, and it is obvious that the ratio of regions of audience is smaller than others. XTable 5.11X shows the saved bits of unstructured regions by our model.
Saved bits in unstructured regions Saved bits 43PrdP 80PthP
Audience 92 105
Table 5.11 Saved bits in unstructured regions in Stefan sequence
Next, in whole Football sequence, human observers may pay more attention on the area of football and football player than the area of grass. Moreover, obvious edges on football player such as numbers on their sports coats or stripes on their pants may attract human eyes dramatically. In XFigure 5.5X(a), the major movement in the 65PthP frame is the football players competing for the football. In the proposed scheme, the performances of the regions we said above that humans may be more sensitive to, numbers on their sports coats or stripes on their pants, are mostly enhanced. Nevertheless, bits which are allocated to the regions of too complicated grass and the flat regions are usually decreased because these regions may be seemed to unstructured regions. And the processing may cause worse visual quality of these regions. Here we select the other kind of frame in Football sequence to analyze its result. In
XFigure 5.5X(b), the major movement in the 120PthP frame is the football players running towards the right with the football in his hand. In this frame, human may pay attention to the only
football player and the football. In our proposed scheme, the performances of the regions we said above mostly enhanced. However, for less important regions, such as the regions of too complicated grass and flat regions, their bits are usually decreased and their visual quality may be reduced.
(a) (b)
Figure 5.5 Comparison of visual quality in Football Sequence (a)The 65PthP frame in Football. (b) The 120PthP frame in Football.
XTable 5.12X shows ratio of number of regions with better PSNR and worse PSNR, and it is obvious that the ratio of regions of grass is smaller than others. XTable 5.13X shows the saved bits of unstructured regions by our model.
65PthP frame 120PthP frame
PSNR Better Worse Better Worse
Grass 79 16 214 21
Numbers 14 1 - -
Stripe 13 1 - -
Whole body - - 31 10
Table 5.12 Ratio of number of regions with better PSNR and worse PSNR in Football sequence
Saved bits in unstructured regions Saved bits 65PthP 120PthP
Grass 101 66
Table 5.13 Saved bits in unstructured regions in Football sequence
Last, in Bus sequence, human observers may not pay more attention on the area of complicated background such as the trees on the top of image, and complicated foreground such as railings and still car. On the other hand, the moving bus and various backgrounds are more attracted to human eyes than the region we said above generally. Here we select two different kinds of scenes of bus sequence to analyze our result. In XFigure 5.6X(a), the bus is just passing through the pillar with sculpture. Therefore, human observers may take their on the regions of the sculpture, human under the sculpture, the head of bus and the top of bus. In our proposed scheme, visual qualities of these regions are mostly enhanced. In the 98PthP frame, as XFigure 5.6X(b) shows, the regions that human observer may notice a lot are listed as follows:
the advertisement with photograph and words on the bus, the street light near the head of bus and the region that sky and trees are associated with. Visual qualities of these regions are mostly enhanced. Nevertheless, for the regions of complicated edges, such as trees, railings and still car, human observers often skip their detail. In our scheme, these regions may be considered unstructured region, and their visual quality may be decrease to save bits.
(a) (b) Figure 5.6 Comparison of visual quality in Bus Sequence
(a)The 47PthP frame in Bus. (b) The 98PthP frame in Bus.
47PthP frame 98PthP frame
PSNR Better Worse Better Worse
Trees 64 32 80 55
Railings 88 70 62 70
Sculpture 9 1 - -
Passerby 2 0 - -
Photo and
Word 8 2 13 2
Table 5.14 Ratio of number of regions with better PSNR and worse PSNR in Stefan sequence
Saved bits in unstructured regions Saved bits 47PthP 98PthP
Tree 92 61 Railings 70 787
Table 5.15 Saved bits in unstructured regions in Bus sequence
XTable 5.14X shows ratio of number of regions with better PSNR and worse PSNR, and it is obvious that the ratio of regions of grass and railings are smaller than others. XTable 5.15X shows the saved bits of unstructured regions by our model.