Multiple reference frames can be adopted to increase the coding efficiency and sup-press the error propagation. However, this technique requires more computational time for mode decision and, especially, for motion estimation, which comsumes most of com-putational time. In general, the comcom-putational complexity depends on the number of candidate reference frames K. Therefore, we propose a fast motion estimation algorithm for the proposed error resilient MRF scheme in order to reduce the computational com-plexity.
5.1 Reference Trend Decision
From the experimental results in figure 3.1, it can be seen that
• more than 90% of blocks choose to reference to the first four reference frames as well as the DIF, and
• about 70% to 80% of blocks reference to the two key frames (the immediately previous frame & the DIF)
5.2 Fast Motion Estimation Algorithm
Thus, we propose to use five reference frames at most, and use two key frames first to decide the trend of prediction because of the different properties of the four previous frames and the DIF. Reference to DIF has good error resilience but poor coding efficiency, while reference to the four previous frames has good coding efficiency but poor error resilience.
Assume current frame is n. With two key frames as reference, if frame n − 1 is selected by proposed ER-RDO formula, it means the gain from coding efficiency is more important than that from error resilience. This could be due to good channel conditions (i.e., low packet loss rate) or video content with high motion. In this case, motion estimation will continue with candidate reference frames n − 2, n − 3 and n − 4. On the contrary, selecting dominant intra frame as reference frame means that error resilience capability is more important than coding efficiency. In this case, candidate reference frames n − 2, n − 3 and n − 4 will be skipped, and motion estimation process can be early terminated.
5.2 Fast Motion Estimation Algorithm
Figure 5.1 shows the H.264 standard ME algorithm.
Step 1. Get next available MB in current frame.
Step 2. Choose next unprocessed mode.
Step 3. Do motion estimation for each candidate reference frame (Fn−i, 1 ≤ i ≤ 5) in selected mode.
Step 4. Apply H.264 standard RDO to ME for each candidate reference frame.
Step 5. If all modes for current MB are done, go to Step 6. Otherwise, go to Step 2.
Step 6. Apply H.264 standard RDO to mode decision to determine the final coding mode.
5.2 Fast Motion Estimation Algorithm
Get next available MB in current frame
All MBs in current frame are done?
Do ME for each candidate frames (Fn-1, Fn-2, Fn-3, Fn-4, Fn-5)
No
End of encoding in current frame
Figure 5.1 H.264 Motion Estimation flow chart
Step 7. If all MBs in current frame are done, go to Step 8. Otherwise, go to Step 1 to get next MB for encoding.
Step 8. End of encoding in current frame.
As described above, the proposed fast motion estimation algorithm changes the order of ME for two key frames as shown in figure 5.2.
Step 1. Get next available MB in current frame.
Step 2. Choose next unprocessed mode.
5.2 Fast Motion Estimation Algorithm
Get next available MB in current frame
Is DIF selected as reference frame?
All MBs in current frame are done?
End of encoding in current frame
Yes
Scene Change? Yes DIF replacement
No
Figure 5.2 Fast Motion Estimation flow chart
Step 3. Do motion estimation for 2 key frames (Fn−1 and FDIF) in selected mode.
Step 4. Apply Error Resilient RDO to ME for 2 key frames.
Step 5. If DIF is selected as reference frame rather than the immediately previous frame,
5.2 Fast Motion Estimation Algorithm
use it as best reference frame in selected mode and go to Step 8. Otherwise, go to Step 6.
Step 6. Do ME for the remaining candidate frames. (Fn−2, Fn−3, Fn−4)
Step 7. Apply Error Resilient RDO to ME for the remaining candidate frames.
Step 8. If all modes for current MB are done, go to Step 9. Otherwise, go to Step 2.
Step 9. Apply Error Resilient RDO to mode decision to determine the final coding mode.
Step 10. If all MBs in current frame are done, go to Step 11. Otherwise, go to Step 1 to get next MB for encoding.
Step 11. Detect whether scene change happened by the ratio of intra-coded MBs in current frame. If scene change happened, go to Step 12. Otherwise, go to Step 13.
Step 12. Apply DIF replacement mechanism.
Step 13. End of encoding in current frame.
With the comparison between figure 5.1 and figure 5.2, it can be seen that the proposed fast motion estimation algorithm, which only changes the order of ME for two key frames without affecting to the ER-RDO applying to both motion estimation and mode decision stages.
5.3 Motion Vector Prediction for DIF
5.3 Motion Vector Prediction for DIF
Since the dominant intra frame may be far from current frame, using co-located MB as the center of search window in motion estimation may not be adequate. Thus, we have designed a scheme to predict MV on the DIF. Our approach is based on FDVS (Forward Dominant Vector Selection) [17], which is a MV composition method originally designed for transcoding with frame skipping applications.
F
nF
n-1F
n-2MV(n, n-1) MV(n-1, n-2)
MV
offset(n, n-1)
MV(n, n-2)
Figure 5.3 FDVS - Fast Dominant Vector Selection
Considering a simple sample as shown in figure 5.3, we have motion M V (n, n − 1) between frame n and frame n − 1, and M V (n − 1, n − 2) between frame n − 1 and frame n − 2, respectively. If frame n − 1 is skipped, the most common way to represent the corresponding motion between frame n and frame n − 2 would be M V (n, n − 1) + M V (n − 1, n−2). However, for each MB in frame n, the area pointed by its MV may not be aligned on the MB boundary of frame n − 1. In FDVS method, the MV associated with the MB with the largest overlapping area out of the four neighboring MBs on frame n − 1 is selected. To compensate the shift in frame n − 1, the authors in [18] suggested to add the offset (denoted by M Vof f set(n, n − 1)) to the composed MV. Therefore, the MV from frame n to frame n − 2 can be predicted as
P M V (n, n − 2) = M V (n, n − 1) + M V (n − 1, n − 2) + M Vof f set(n, n − 1) (5.1)
5.3 Motion Vector Prediction for DIF
This MV prediction technique can be easily extended to across more frames. Assume current frame is frame n and DIF is frame m (m < n). The predict MV from frames n to m can be represented as
P M V (n, m) =
n−1
X
i=m
[M V (i + 1, i) + M Vof f set(i + 1, i)] (5.2)
It can be seen that such MV prediction relies on all the MVs between frame n and frame m. However, all the required MVs may not be available in the encoding buffer, especially when n−m > k, where k is the number of reference frames. To rope with this problem, we propose a modified FDVS scheme called Accumulated Forward Dominant Vector Selection (AFDVS) by using Accumulated Motion Vector (AMV).
First of all, we define AM V (i) to denote the predict MV from frame i to DIF P M V (i, DIF ). Assume each frame i has only one candidate reference frame, which is its immediately previous frame i − 1. Therefore, by equation (5.1), AM V (i) can be recursively derived from AM V (i − 1) as
AM V (i) = M V (i, i − 1) + AM V (i − 1) + M Vof f set(i, i − 1) (5.3)
where the initial AM V (m) = 0 if DIF is frame m. From (5.3), it is interesting to notice that once AM V (i) is obtained, AM V (i − 1), M V (i, i − 1) and M Vof f set(i, i − 1) are unnecessary for calculating AM V (i + 1), that is, they can be removed from the encoding buffer. In other words, if AM V in reference frame (frame n − 1) is available, the predict MV from current frame n to DIF (frame m) can be derived by
P M V (n, m) = M V (n, n − 1) + AM V (n − 1) + M Vof f set(n, n − 1) (5.4)
To be more general, we further modify the equation (5.4) for MRF, as shown in figure
5.3 Motion Vector Prediction for DIF
F
nF
refF
m(DIF)...
MV(n, ref) AMV(ref)
MVoffset(n, ref) PMV(n, m)
Figure 5.4 AFDVS - Accumulated Fast Dominant Vector Selection
as
P M V (n, m) = M V (n, ref ) + AM V (ref ) + M Vof f set(n, ref ) (5.5)
where ref denotes the frame number of reference frame.
The equation (5.5) shows that the current predict MV can be calculated by tracing the reference frame. Since the MV prediction may across all candidate reference frames, even if AM V (n) is obtained, AM V (ref ) still cannot be immediately removed from the encoding buffer.
To be more clearly, assume the number of reference frames K is 3, as shown in figure 5.5. At frame 1, it only predicts from DIF (frame 0), thus the predict MV P M V (1, 0) is simply equal to M V (1, 0). At frame 2, it may predict from previous frame 1 or from DIF, thus these two cases has to be considered to estimate the predict MV P M V (2, 0).
If it chooses frame 0 as reference frame, the MV M V (2, 0) is directly set as P M V (2, 0), since AM V (0) is 0, otherwise, it combines M V (2, 1) and AM V (1), which is calculated at frame 1, as P M V (2, 0). At frames 3 to 5, there are only 3 cases needs to be considered (K = 3), and by using the same way at frame 2, the predict MVs P M V (3, 0), P M V (4, 0) and P M V (5, 0) can be easily composed. Thus, a general formula is proposed to represent
5.4 Time Complexity of Fast Motion Estimation
the accumulated predict MV P M V (n, 0) from frames n to DIF
P M V (n, 0) =
From equation (5.6), it can be seen that, to obtain predict MV P M V (n, 0) from current frame n to DIF, we only need 2 AM V s from frame n − 1 and from frame n − 2 when K = 3. That is, at most of K − 1 AM V s need to be keep in encoding buffer.
Moreover, an accurate predict MV to DIF can also shrink the search range of motion estimation to reduce the computational complexity. In our thesis, the search range on DIF is set as 4.
5.4 Time Complexity of Fast Motion Estimation
Assume the following
• number of reference frames: k
• search range of DIF: from 32 to 4
• reference ratio of DIF (skip ME of k − 2 reference frames): r (about 14% to 26%)
Thus, the time complexity ratio of ME is estimated as
r 2
5.4 Time Complexity of Fast Motion Estimation
By applying fast motion estimation algorithm to proposed scheme, the computational complexity in motion estimation can be substantially reduced.
5.4 Time Complexity of Fast Motion Estimation
PMV(1,0)
F
0(DIF)F
1AMV(1)
F
0(DIF)F
1F
2MV(2,0) MV(2,1) PMV(2,0)
AMV(1)
F
0(DIF)F
1F
2F
3MV(3,0)
MV(3,1) AMV(2)
MV(3,2) PMV(3,0)
MV(1,0)
AMV(2)
F
0(DIF)F
1F
2F
3AMV(3)
AMV(4)
F
0(DIF)F
1F
2F
3F
4F
5MV(5,0)
MV(5,3)
MV(5,4)
AMV(3) PMV(5,0)
MV(4,2) MV(4,0)
MV(4,3)
F
4PMV(4,0)
Chapter 6
Experimental Results
The proposed ER-RDO-based MRF method is integrated into the latest JVT reference software JM15.1 [19]. The parameters of our experimental environment are set as follows.
• Test sequence: Foreman, Football, News
• Number of frames: 100 frames
• Frame rate: 30 fps
• Structure of the Group of Picture (GOP): I P P P P ...
• GOP size: 30 frames
• Frame format: CIF (352 × 288 pixels)
• Number of slices per frame: 6 slices
• 100 random loss patterns for each of different packet loss rates (1%, 5%, 10%)
The methods used for comparison are listed as follows.
• H.264: the original RDO method in H.264 standard.
6.1 Bit Rate v.s. Average PSNR
• Periodic MB: Periodic Macroblock proposed by J. Zheng in [13], where the period K = 5.
• Robust MB: Robust Macroblock proposed by Q. Zhang in [14], where the number of robust MBs per frame is 120.
• ER-RDO: The ER-RDO model in [12]. Note that, in [12], the ER-RDO is applied to mode decision stage only.
• ER-RDO [12] + ME: The ER-RDO model in [12] is used, but it is applied to both the mode decision stage and the motion estimation stage.
• Proposed w/o Fast ME: the proposed method without fast motion estimation mechanism.
• Proposed w/o IMB: the proposed method without end-to-end distortion estima-tion for intra-coded MBs.
• Proposed: the full version of proposed method.
6.1 Bit Rate v.s. Average PSNR
The measured average PSNR results of sequences Foreman, Football, News with packet loss rates 1%, 5% and 10% are shown in figure 6.1, figure 6.2, figure 6.3 and table 6.1.
We can see that the proposed method is better than methods proposed in [12–14]. It is observed that for sequence Foreman in figure 6.1, the Robust MB scheme in [14] achieves good performance at packet loss rates 1% and 5%, however, it has about 0.7 dB lower than the proposed method at packet loss rate 10%; And for sequence News in figure 6.3,
6.1 Bit Rate v.s. Average PSNR
0.8 dB lower than the proposed method at packet loss rate 1%. The results are due to that although the scheme in [14] choosed Robust MBs according to estimated distortion, the number of Robust MBs for each frame is a fixed constant which makes it hard to adopt to different channel conditions. On the other hand, the scheme in [14] achieved good performance for sequences Foreman and News, but it has about 2.5 dB lower than the proposed method for sequence Football. The results again show that a fixed number of Robust MBs per frame is not suitable for various sequences. The proposed method adopts ER-RDO at the stage of ME, which will select the best MV and reference frame for each MB, resulting in an optimal number of MBs that reference to intra-coded MBs.
Therefore, the proposed method is adaptive to varying channel conditions and various video sequences.
According to the results of ‘Proposed w/o IMB’ and ‘ER-RDO [12]’, by involving DIF as one of reference frames, ‘Proposed w/o IMB’ can gain about 1.2 dB higher than
‘ER-RDO [12]’ for sequence Foreman and about 0.7 dB higher for sequence News. The reason that the gain for sequence Football is relatively low (less than 0.5 dB) is due to that Football is a high-motion sequence and, referencing to DIF will cause too much increase in the coding bits. As a consequence, a relative high ratio of MBs choose intra-coding to alleviate error propagation. This makes the performance difference between
‘Proposed w/o IMB’ and ‘ER-RDO [12]’ become small. By involving end-to-end distortion estimation for intra-coded MBs, the proposed method can still gains more than 0.7 dB in low packet loss rate, and especially, more than 2 dB in high packet loss rate.
6.1 Bit Rate v.s. Average PSNR
Loss Rate 1% 5% 10%
QP 30 29 28 27 26 30 29 28 27 26 30 29 28 27 26
Foreman
Proposed (5 Ref) BR 478.54 558.38 672.85 801.63 925.52 615.96 726.59 871.34 1048.37 1226.96 741.82 876.60 1048.78 1250.61 1445.71
PSNR 33.69 34.17 34.73 35.24 35.60 30.63 30.87 31.52 31.88 32.20 28.48 28.78 29.45 29.69 29.84
H.264 BR 420.51 488.40 581.65 688.73 788.36 420.51 488.40 581.65 688.73 788.36 420.51 488.40 581.65 688.73 788.36
PSNR 32.15 32.52 32.72 33.27 33.38 27.94 28.08 28.33 28.32 28.42 25.18 25.20 25.35 25.39 25.41
Periodic MB [13]
(Period 5)
BR 503.16 590.63 710.95 847.18 980.75 503.16 590.63 710.95 847.18 980.75 503.16 590.63 710.95 847.18 980.75
PSNR 33.54 34.01 34.56 34.96 35.42 29.47 29.62 29.83 29.92 30.15 26.54 26.58 26.70 26.71 26.81
Robust MB [14] BR 595.12 682.82 820.31 962.12 1104.80 595.12 682.82 820.31 962.12 1104.80 595.12 682.82 820.31 962.12 1104.80
PSNR 34.02 34.52 35.18 35.68 36.22 30.63 30.97 31.33 31.60 31.86 27.78 28.03 28.22 28.39 28.57
ER-RDO [12] (5 Ref)
BR 459.01 534.34 642.54 758.73 878.86 555.06 650.45 777.01 926.23 1072.47 627.80 731.27 869.33 1027.60 1185.56
PSNR 32.26 32.86 33.11 33.49 33.83 29.11 29.49 29.61 30.15 30.33 26.77 27.06 27.24 27.50 27.59
[12] + ME (5 Ref)
BR 464.06 540.69 644.44 764.82 884.90 567.85 666.89 800.34 946.59 1090.68 652.83 762.64 908.26 1066.61 1229.08
PSNR 32.25 32.69 32.82 33.48 33.67 29.19 29.56 29.75 30.26 30.16 26.84 26.85 27.08 27.58 27.58
Football
Proposed (5 Ref) BR 1561.42 1742.15 1990.05 2246.20 2475.27 1743.87 1924.36 2187.41 2450.12 2678.87 1819.74 2002.47 2270.71 2541.25 2775.20
PSNR 34.00 34.56 34.87 35.79 36.65 30.78 31.46 31.98 32.66 33.23 28.77 29.32 29.80 30.25 30.94
H.264 BR 1413.31 1574.89 1796.35 2025.08 2216.43 1413.31 1574.89 1796.35 2025.08 2216.43 1413.31 1574.89 1796.35 2025.08 2216.43
PSNR 32.16 32.87 33.28 33.74 33.92 26.20 26.37 26.44 26.87 26.45 23.07 23.24 23.19 23.44 23.12
Periodic MB [13]
(Period 5)
BR 1452.26 1614.98 1840.68 2073.20 2271.72 1452.26 1614.98 1840.68 2073.20 2271.72 1452.26 1614.98 1840.68 2073.20 2271.72
PSNR 33.10 33.45 34.18 34.63 35.19 27.01 27.19 27.52 27.46 27.77 23.66 23.81 24.09 24.01 24.26
Robust MB [14] BR 1590.18 1751.79 1987.37 2222.81 2433.85 1590.18 1751.79 1987.37 2222.81 2433.85 1590.18 1751.79 1987.37 2222.81 2433.85
PSNR 33.86 34.23 34.92 35.47 36.04 28.53 28.65 28.99 29.27 29.66 25.22 25.09 25.42 25.64 25.96
ER-RDO [12] (5 Ref)
BR 1491.89 1662.10 1896.45 2137.11 2348.33 1566.64 1738.76 1978.30 2226.08 2437.10 1598.62 1769.74 2014.96 2261.93 2476.05
PSNR 32.64 33.21 34.12 34.71 35.22 27.82 28.15 28.50 29.11 29.22 24.66 24.92 24.94 25.32 25.36
[12] + ME (5 Ref)
BR 1497.06 1664.85 1901.07 2138.51 2349.88 1586.71 1757.63 2003.59 2247.24 2459.89 1636.38 1808.78 2056.69 2307.52 2516.06
PSNR 32.79 33.29 34.07 34.58 35.05 27.93 28.30 28.54 28.89 29.08 25.01 25.22 25.58 25.62 25.68
News
Proposed (5 Ref) BR 275.90 309.95 355.23 402.50 449.78 307.47 348.30 403.25 460.43 520.42 347.76 392.22 454.35 520.36 586.65
PSNR 35.68 36.33 36.97 37.58 38.15 32.39 32.82 33.37 33.78 34.19 29.73 30.10 30.56 30.87 31.24
H.264 BR 266.02 298.40 340.54 385.44 428.55 266.02 298.40 340.54 385.44 428.55 266.02 298.40 340.54 385.44 428.55
PSNR 35.13 35.82 36.41 36.94 37.42 30.70 31.12 31.35 31.52 31.79 27.67 27.91 28.04 28.12 28.25
Periodic MB [13]
(Period 5)
BR 291.02 326.38 372.74 419.35 465.54 291.02 326.38 372.74 419.35 465.54 291.02 326.38 372.74 419.35 465.54
PSNR 35.78 36.36 37.04 37.61 38.21 32.02 32.33 32.79 33.11 33.41 28.97 29.10 29.42 29.64 29.84
Robust MB [14] BR 338.89 379.21 430.13 484.33 539.83 338.89 379.21 430.13 484.33 539.83 338.89 379.21 430.13 484.33 539.83
PSNR 35.77 36.43 37.04 37.66 38.21 32.56 33.04 33.44 33.82 34.18 29.74 30.07 30.28 30.56 30.85
ER-RDO [12] (5 Ref)
BR 271.43 304.92 350.41 397.48 442.27 291.01 327.16 377.02 430.54 483.54 309.66 347.49 401.18 458.72 513.13
PSNR 35.38 35.93 36.64 37.22 37.59 31.45 31.82 32.30 32.62 32.96 28.58 28.97 29.27 29.59 29.54
[12] + ME (5 Ref)
BR 271.75 305.00 349.76 396.54 442.50 293.81 330.49 380.03 432.59 483.60 314.31 354.39 407.67 462.49 514.38
PSNR 35.28 35.99 36.66 37.17 37.57 31.48 31.94 32.28 32.74 33.09 28.63 29.07 29.30 29.63 29.66
Table 6.1 Table of Bit Rate v.s. Avg. PSNR in different packet loss rates p = 0.01, 0.05, 0.10 and QP = 26, 27, 28, 29, 30
6.2 The Effects of Intra-coded MBs in End-to-End Distortion Estimation
Sequence Method Percentage of intra-coded MBs p = 1% p = 5% p = 10%
Foreman Proposed (5 Ref) 10% 22% 32%
Proposed w/o IMB (5 Ref) 8% 16% 20%
Football Proposed (5 Ref) 53% 72% 80%
Proposed w/o IMB (5 Ref) 42% 49% 53%
News Proposed (5 Ref) 3% 6% 8%
Proposed w/o IMB (5 Ref) 3% 4% 5%
Table 6.2 Intra-coded MB Rates with different packet loss rates p = 0.01, 0.05, 0.10 and QP = 28 for sequences Foreman, Football and News
6.2 The Effects of Intra-coded MBs in End-to-End Distortion Estimation
As described in chapter 4, our proposed method considers intra-coded MBs in end-to-end distortion estimation to get accurate expectation of end-to-end-to-end-to-end distortion. Therefore, in our experiment, we compare methods ‘Proposed’ with ‘Proposed w/o IMB’. Table 6.2 shows the intra-coded MB rate for different sequences and different packet loss rates with QP = 28. It can be seen that, the inclusion of intra-coded MBs for end-to-end distortion causes increase of the intra-coded MB rate. That is because, actually, the intra-coded MBs in inter frames may cause error propagation if it is not received. Without considering it, the end-to-end distortion estimation may underestimate the impact of error propagation from intra-coded MBs, especially for high-motion equences and high packet loss rate, since they usually reveals in high intra rate. Figure 6.4 shows the average PSNR frame by frame with packet loss rate 10% and QP = 28 for different sequences Foreman, Football and News. Since the intra-coded MB rate has a big gap (27%) between the methods with and without consideration of intra-coded MB in end-to-end distortion estimation for sequence Football with packet loss rate 10%, the degradation of average PSNR is high.
6.3 Performance of Fast Motion Estimation
6.3 Performance of Fast Motion Estimation
The proposed fast motion estimation algorithm (FME) involves AFDVS to predict MVs on DIF rather than the way using co-located MBs. Moreover, two key frames (the DIF and the immediately previous frame) are used to determine the trend of prediction.
Table 6.3 shows the reference ratios referencing to DIF and the four previous frames.
From the table, we can see that, by involving fast motion estimation algorithm, the percentage of reference to the DIF increases about 4.2%, 1.2% and 0.3% at low packet loss rate for Foreman, Football and News, respectively. At high packet loss rate, the difference of reference ratio is about 6.1%, 1.8% and 0.6% for Foreman, Football and News, respectively. Those blocks which change reference frame to DIF may provide more resilience against error propagation. However, referencing to the DIF may require more coding bits. From the performance comparison, as shown in figure 6.5, the performance of proposed method with FME is better than the one without FME for News and, especially, for Foreman, however, it’s worse for Football. The main reason is that, for low intra rate sequences such as Foreman and News, the gain of referencing to the DIF is higher enough to counterbalance against the increase of coding bits. The effect is obvious at high packet loss rate. However, for high-motion sequences such as Football, the gain of referencing to the DIF is quite small due to high intra rate. Futhermore, the coding bits by referencing to the DIF is much higher for high-motion sequences than the one for low-motion sequences.
That is the reason that, for Football, the proposed method with FME gains lower than the method without FME. Even though fast motion estimation causes lower gain for high-motion sequences, it still costs low computational complexity, which is described in chapter 5.
6.4 Performance of DIF Replacement Mechanism
Sequence Method Reference Frame Percentage of Reference
p = 1% p = 5% p = 10%
Foreman
Proposed (5 Ref) DIF 14.2% 21.4% 28.1%
Others 85.8% 78.6% 71.9%
Proposed w/o Fast ME (5 Ref) DIF 10.0% 16.4% 22.0%
Others 90.0% 83.6% 78.0%
Football
Proposed (5 Ref) DIF 10.3% 19.2% 27.0%
Others 89.7% 80.8% 73.0%
Proposed w/o Fast ME (5 Ref) DIF 9.1% 16.4% 25.2%
Others 90.8% 83.6% 74.8%
News
Proposed (5 Ref) DIF 5.3% 5.7% 6.2%
Others 94.7% 94.3% 93.8%
Proposed w/o Fast ME (5 Ref) DIF 5.0% 5.2% 5.6%
Others 95.0% 94.8% 94.4%
Table 6.3 Reference Ratios with different packet loss rates p = 0.01, 0.05, 0.10 and QP = 28 for sequences Foreman, Football and News
6.4 Performance of DIF Replacement Mechanism
In chapter 4, we propose a DIF replacement mechanism which adaptively changes the dominant intra frame when scene change happened. In the experiment, we cascade sequences Foreman and Stefan as a composite sequence of 100 frames with scene changes at frame 34 and frame 63. Figure 6.6 shows the performance comparison of the proposed methods with and without DIF replacement mechanism. According to the results, we can see that, by involving DIF replacement mechanism, the performance gain is about 0.1 dB and 0.3 dB for packet loss rates low and high, respectively.
6.4 Performance of DIF Replacement Mechanism Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5)
6.4 Performance of DIF Replacement Mechanism Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5)
Figure 6.2 Football - Bit Rate v.s. Average PSNR with p = 0.01, 0.05, 0.10
6.4 Performance of DIF Replacement Mechanism Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5) Proposed w/o IMB (Ref 5) Proposed (Ref 5)
6.4 Performance of DIF Replacement Mechanism
Figure 6.4 Frame v.s. PSNR with p = 0.10 for sequences (a) Foreman (b)
6.4 Performance of DIF Replacement Mechanism Proposed (Ref 5, LR 1%) Proposed w/o Fast ME (Ref 5, LR 5%) Proposed (Ref 5, LR 5%) Proposed w/o Fast ME (Ref 5, LR 10%) Proposed (Ref 5, LR 10%)
28.5 Proposed (Ref 5, LR 1%) Proposed w/o Fast ME (Ref 5, LR 5%) Proposed (Ref 5, LR 5%) Proposed w/o Fast ME (Ref 5, LR 10%) Proposed (Ref 5, LR 10%)
29.5 Proposed (Ref 5, LR 1%) Proposed w/o Fast ME (Ref 5, LR 5%) Proposed (Ref 5, LR 5%) Proposed w/o Fast ME (Ref 5, LR 10%) Proposed (Ref 5, LR 10%)
6.4 Performance of DIF Replacement Mechanism Proposed w/o DIF Replacement (Ref 5)
30 Proposed w/o DIF Replacement (Ref 5)
28.5 Proposed w/o DIF Replacement (Ref 5)
Figure 6.6 Foreman-Stefan - Bit Rate v.s. Average PSNR with p =
Chapter 7
Conclusion
In this thesis, an RDO-based error resilient scheme using MRF has been presented.
We propose a candidate reference frame set with inclusion of the dominant intra frame (called DIF) to suppress the error propagation and an error resilient RDO model applied to both stages Motion Estimation and Mode Decision for better coding efficiency in dif-ferent network conditions and difdif-ferent contents. Futhermore, we propose a fast motion estimation algorithm to reduce the computational complexity. To fit in the situation with scene changes, we also propose a replacement mechanism for DIF. The experimental re-sults show that, our proposed scheme improves coding efficiency with different content and network conditions but without much increase of computational complexity in the MRF environment.
References
[1] J. Mochnac and S. Marchevsky, “Error resilience tools in the mpeg-4 and h.264 video
[1] J. Mochnac and S. Marchevsky, “Error resilience tools in the mpeg-4 and h.264 video