• 沒有找到結果。

Combined Data Efficient INTER and IL Prediction Algorithm

Chapter 3. Proposed Data Efficient Algorithm for INTER and ILR

4.3. Combined Data Efficient INTER and IL Prediction Algorithm

Prediction Algorithm

From previous description, we proposed a data efficient ILM and an Inter-BL prediction algorithm to solve the problem of bandwidth overhead and computational complexity. However, the data access requirement of INTER prediction has also to be discussed in this thesis since all our proposed data efficient IL algorithms have reused the reference data of INTER prediction. For reducing the data access requirement, the level C data reuse scheme has been widely adopted in designing motion estimator. Nevertheless, the data access requirement of level C data reuse scheme becomes noticeable when search range larger than ±64 for SD/HD sequences. Therefore, the reference data subsample motion estimation algorithm called PMRME was proposed to solve the reference data access intensive problem. As a result, we analyze the data access requirement for these two methods (level C data reuse scheme and PMRME) and choose the best one which can achieve least data access requirement when combined with our proposed inter-layer prediction

43

algorithms. Table 4.3 shows the analysis results of data access requirements, and the statistic chart is shown in Fig. 4.7. The IL data reuse denotes the whether the reference data of INTER is reused in IL mode. The Level C data reuse scheme with IL data reuse denotes that the Level C data reuse scheme is applied for INTER mode to load reference data and IL prediction reuses INTER mode’s reference data. Otherwise, the reference data of IL centered at MVPILM is loaded from external memory without data reuse. When the PMRME with IL data reuse is adopted, it represents the [-8, +7] reference data of INTER mode is reused. Otherwise, the reference data of all IL prediction modes are loaded from external memory without data reuse. To future compare the data access with different cases, the percentage of data access requirement is shown in Fig. 4.7. The result shows that when the resolution is smaller or equal to CIF size, level C data reuse scheme for both INTER and IL prediction is preferred. Otherwise, when the resolution is larger or equal to 480p size, the PMRME with IL data reuse algorithm is a better choice.

Table 4.3. Data access requirement in different frame resolutions (Kbyte/frame)

1 Compared to Method 1

IL data reuse QCIF CIF 480p 4CIF 720p 1080p

Level C N (Method 1) 173.77 1300.16 10626.66 12472.60 83592.05 636766.44 Y (Method 2) 78.63 425.39 2201.31 2586.85 9975.65 41894.34

PMRME N (Method 3) 215.62 1435.77 10525.77 12351.20 80243.69 617101.49 Y (Method 4) 115.67 541.74 2034.76 2388.41 6452.18 21835.40

1Data access saving (%) 54.74 67.28 80.85 80.85 92.28 96.57

44

QCIF CIF 480p 4CIF 720p 1080p

Resolution

Log scale (Kbyte/frame)

Method 1 Method 2 Method 3 Method 4

Fig. 4.7.The data access requirement in different frame resolutions

Data access requirement

QCIF CIF 480p 4CIF 720p 1080p

Resolution

Percentage

Method 1 Method 2 Method 3 Method 4

Fig. 4.8.The percentage of data access requirement in different frame resolutions

45

Therefore, we choose Level C data reuse scheme for both INTER and proposed IL data reuse algorithms when the resolution is smaller than or equal to CIF size. Otherwise, the PMRME with proposed IL data reuse algorithms is selected as proposed algorithm.

The combination of proposed algorithm is shown in Fig. 4.9. In the first step, the INTER prediction process is executed and the reference data corresponding to different ME algorithm would be loaded into internal memory.

Afterward, the reference data of INTER prediction can be reused during the ILM and Inter-BL process. Finally, the best mode is selected by comparing the

RD_Cost. The detailed flow charts of these three modules are described

below.

end start

Frame adaptive ME switching method for

INTER prediction

Proposed data reuse Inter-BL algorithm

Obtain the RD_Cost for

Inter-BL

Obtain the best mode

Fig. 4.9.The flowchart of combined data efficient IL algorithms and dynamic ME switching method

46

Fig. 4.10 is our proposed frame size adaptive ME switch algorithm. The ME algorithm is determined by the frame resolution. When the frame resolution is larger than or equal to 480p, we choose the PMRME algorithm for the reason that the data access requirement of PMRME is less than Level C scheme. Otherwise, the Level C scheme is selected in ME operation.

Obtain the RD_Cost for INTER mode

Derive the MVPINTER

INTER prediction

Load the reference data by

Level C scheme start

Derive the MVPINTER

Load the reference data for PMRME algorithm

end

Y N

Frame resolution

>= 480p

Fig. 4.10.The flowchart of frame size adaptive ME switching method

As mentioned in the previous section, the ILM mode is able to reuse the reference data of INTER mode. Since the INTER prediction adopts frame size adaptive ME switching method, the reference data is distinct when choosing different ME algorithms. Therefore, the data efficient ILM algorithm is altered rely on the ME algorithm. Fig. 4.11 shows flowchart of the combined data efficient ILM algorithm and adaptive ME switching algorithm. When the frame

47

resolution is larger than or equal to 480p, the PMRME is used in INTER process. The difference between MVPILM and MVPINTER is checked to determine whether reference data of INTER prediction with [-8, +7] search range size would be reused by ILM mode or not. When the Level C data reuse scheme is selected in INTER process, the entire reference data existed in internal memory can be efficiently used for ILM mode.

Reuse reference data of INTER prediction for ILM

with SR [-8, +7]

Obtain the RD_Cost for ILM Derive the MVPILM

start

Y N

Derive the MVPILM

Reuse reference data of INTER prediction for ILM

with Level C scheme

end

Fig. 4.11.The flowchart of combined data efficient ILM algorithm and adaptive ME switching algorithm

For Inter-BL mode, the data efficient algorithm is adopted with a motion detection mechanism. The detection is based on a threshold which related to the search range of reference data. The flowchart of combined data efficient Inter-BL algorithm and adaptive ME switching algorithm is shown in Fig. 4.12.

48

When PMRME is used, the threshold is set to 8 because the reference data is available within the search window of INTER prediction mode. Otherwise, the threshold is set to the search range size of Level C scheme due to the reference data is existed inside the search range. After setting the threshold, the motion detection mechanism is performed for each partition. When the partition n of MVInterBL_n is less than or equal to the threshold, it implies that the reference data is already existed and can be reused. If the partition n of MVInterBL_n is over the threshold, it is necessary to load the reference data from external memory for Inter-BL prediction.

49

Reuse reference data of INTER prediction

For Inter-BL mode

Derive Mbpartition, n=0

Y

Loaded reference data from external memory

n = n+1 Accumulate the

RD_Cost

N

Derive MVInterBL for partition n; MVInterBL_n

n < MBpartition

threshold = 8,Apply PMRME algorithm

threshold = SR, Apply Level C ME algorithm

Fig. 4.12.The flowchart of combining proposed data reuse Inter-BL algorithm and adaptive ME switching algorithm

4.4. Simulation Results

The proposed algorithms are implemented on a JSVM9.14 [12], and the simulation conditions are shown in Table 4.4.

50

Table 4.4. Simulation conditions

Codec JSVM 9.14 encoder

Test sequences

Akiyo, Coastguard, Container, Foreman, Mobile, Silent, Blue_sky, Tractor, Station2,

Pedestrain_area, Rush_hour, Riverbed

QP 16, 20, 24, 28, 32

Resolution QCIF and CIF

CIF and 480p

IV. INTER prediction search algorithm with Level C and PMRME

V. Proposed IL prediction algorithm in enhancement layer

Fig. 4.13 and Fig. 4.14 show the RD performance in CIF and 480p size sequences, respectively. Table 4.5 and Table 4.6 show the detailed RD performance comparison. The simulation results show that, when compared to JSVM 9.14, the RD performance is 0.16% in bit-rate increase and 0.002dB in PSNR degradation for CIF sequences when the search algorithm is full search with Level C data reuse scheme for INTER prediction and reference data reuse for IL prediction, and the RD performance is 0.91% in bit-rate increase and 0.09dB in PSNR degradation in 480p sequences if the PMRME search algorithm with data efficient IL prediction method is applied.

51

0 500 1000 1500 2000 2500 3000 3500

Bit-rate (Kbit/sec)

PSNR

JSVM9.14

LevelC + IL data reuse PMRME + IL data reuse

Foreman_CIF

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Bit-rate (Kbit/sec)

PSNR

JSVM9.14

LevelC + IL data reuse PMRME + IL data reuse

Fig. 4.13.The RD-curve of Level C + IL and PMRME + IL data reuse for CIF sequences (a) Coastguard (b) Foreman

52

Table 4.5.The RD performance comparison of proposed data efficient IL prediction algorithms with different ME approach for CIF sequences.

CIF sequences@15Hz QP

16 20 24 28 32 PMRME average : ΔBitrate = 1.123%, ΔPSNR = -0.073 dB

53

0 1000 2000 3000 4000 5000 6000

Bit-rate (Kbit/sec)

PSNR

JSVM9.14

Level C + IL data reuse PMRME + IL data reuse

(a)

0 2000 4000 6000 8000 10000 12000

Bit-rate (Kbit/sec)

PSNR

JSVM9.14

Level C + IL data reuse PMRME + IL data reuse

(b)

Fig. 4.14.The RD-curve of Level C + IL and PMRME + IL data reuse for 480p sequences (a) pedestrian_area (b) tractor

54

Table 4.6.The RD performance comparison of proposed data efficient IL prediction algorithms with different ME approach for 480p sequences.

480p sequences@15Hz QP

16 20 24 28 32

55

4.5. Summary

In this chapter, several data efficient IL prediction algorithms are proposed to solve the problem of data access demands brought by IL prediction of SVC.

By the high correlation between INTER and IL prediction, we propose data reuse schemes for INTER and ILM prediction. And a motion detection mechanism is proposed to reuse the INTER reference data for Inter-BL mode.

The proposed efficient ILR is also combined to construct a complete data efficient IL algorithm.

In addition, we further propose a frame size adaptive ME switching method. The method dynamically selects the ME algorithm by taking the frame resolution into account with the data access consideration.

Analysis results show that the proposed method can save 67.3% and 80.9% data access requirement for CIF and 480p frame size, respectively. The simulation results show that the performance is 0.16% in bit-rate increase and 0.002dB in PSNR degradation in CIF sequences when the search algorithm is full search with Level C data reuse scheme for both INTER and IL predictions, the performance is 0.91% in bit-rate increase and 0.09dB in PSNR degradation in 480p sequences when the search algorithm is PMRME with data efficient IL prediction method, and when compared to JSVM 9.14.

56

57

相關文件