Academia.eduAcademia.edu

Figure 4 – uploaded by Chang Wen Chen

See full PDFdownloadDownload figure
COMPARISON OF DIFFERENT METHODS WITH HEVC RA TABLE IV where C, and C> are two constants. Hence, we determine O; as the quantization step size that is least different from 01 in logarithmic domain. Qp, is then set as the corresponding Qp of Q). Specially, to avoid quality fluctuation between GOPs, we clip the Qps of all layers to be within [Q p—3, QO p+3l, where Q p is the average Qp of all layers calculated by the static Qp scheme.

Table 4 COMPARISON OF DIFFERENT METHODS WITH HEVC RA TABLE IV where C, and C> are two constants. Hence, we determine O; as the quantization step size that is least different from 01 in logarithmic domain. Qp, is then set as the corresponding Qp of Q). Specially, to avoid quality fluctuation between GOPs, we clip the Qps of all layers to be within [Q p—3, QO p+3l, where Q p is the average Qp of all layers calculated by the static Qp scheme.

Related Figures (14)

PERCENTAGES OF USING THE NEAREST REFERENCES IN HEVC RA TABLE I which is similar to that in [28]. The coding performance of our conclusion will be examined by HEVC default multiple reference settings in Section V.
PERCENTAGES OF USING THE NEAREST REFERENCES IN HEVC RA TABLE I which is similar to that in [28]. The coding performance of our conclusion will be examined by HEVC default multiple reference settings in Section V.
Fig. 1. An example of an RA prediction structure with a GOP size of 8. influenced by its reference frames in a lower layer. This RD dependency was studied in [28] for HEVC default RA setting. An approximately linear inter-layer D dependency was observed as fe: es
Fig. 1. An example of an RA prediction structure with a GOP size of 8. influenced by its reference frames in a lower layer. This RD dependency was studied in [28] for HEVC default RA setting. An approximately linear inter-layer D dependency was observed as fe: es
COMPARISON II OF DIFFERENT D-Q AND R-Q MODELS TABLE III
COMPARISON II OF DIFFERENT D-Q AND R-Q MODELS TABLE III
COMPARISON I OF DIFFERENT D-Q AND R-Q MODELS TABLE II
COMPARISON I OF DIFFERENT D-Q AND R-Q MODELS TABLE II
According to [37], the convergence is quadratic if A), B; and I; are sufficiently smooth and the initial point is close to one of the roots of the equations. In this work, we set xeE(0) = {0;0;...;0; 1} and the iteration is stopped when |J~'£(xg(1)) |? < 0.01 - (L + 1), which yields an average error of AQ) to be approximately less than 0.1. The iteration process usually takes less than 10 times. Note that there exists a very small possibility that J~! does not exist if
According to [37], the convergence is quadratic if A), B; and I; are sufficiently smooth and the initial point is close to one of the roots of the equations. In this work, we set xeE(0) = {0;0;...;0; 1} and the iteration is stopped when |J~'£(xg(1)) |? < 0.01 - (L + 1), which yields an average error of AQ) to be approximately less than 0.1. The iteration process usually takes less than 10 times. Note that there exists a very small possibility that J~! does not exist if
Fig. 2. Examples of Y-PSNR improvements at the same bit rates. Horizontal axis: Bit rate; Vertical axis: Y-PSNR improvement. (a) Traffic in Class A. (b) Kimono in Class B. (c) RaceHorses in Class C. (d) RaceHorses in Class D. (e) BasketballDrillText in Class F.
Fig. 2. Examples of Y-PSNR improvements at the same bit rates. Horizontal axis: Bit rate; Vertical axis: Y-PSNR improvement. (a) Traffic in Class A. (b) Kimono in Class B. (c) RaceHorses in Class C. (d) RaceHorses in Class D. (e) BasketballDrillText in Class F.
COMPUTATIONAL COMPLEXITIES OF NEWTON-RAPHSON METHOD IN TERMS AVENUMITERPERGOP, AVETIMEPERITER (15) AND TOTITERTIMEPPM
COMPUTATIONAL COMPLEXITIES OF NEWTON-RAPHSON METHOD IN TERMS AVENUMITERPERGOP, AVETIMEPERITER (15) AND TOTITERTIMEPPM
CONTRIBUTION OF ADAPTATION IN THE PROPOSED METHOD TABLE VII AveNumlterPerGOP, represents the average number of iterations when perform QPC in one GOP; AveTimePerlter, represents the average time (in terms of ys) per iteration; and TotIterTimePPM, represents the total iteration time in terms of part per million of the overall encoding time. The test was run on a 2011 Laptop with a Quad-Core i17-2.70GHz CPU and 8GB memory. The measure results are summarized in Table V. It can be seen that on average, less than 10 times of Newton-Raphson iterations are required to perform QPC and the average time of each iteration is less than 1 ws. As the Qp increase, the AveNumlIterPerGOP slightly increases. Besides, the total time cost of these iterations is less than | part per mil- lion of the overall encoding time. Therefore, the computation overhead of Newton-Raphson method is negligible. proposed approach is negligible. Investigations of the Newton- Raphson method also show that, less than 10 iterations are required for most cases, which ensures a low computational complexity. Therefore, the coding time differences between the original HEVC and the proposed approaches are mostly determined by the changing of Qps, which may be positive or negative. Therefore, our scheme may even reduce the overall coding time in some cases.
CONTRIBUTION OF ADAPTATION IN THE PROPOSED METHOD TABLE VII AveNumlterPerGOP, represents the average number of iterations when perform QPC in one GOP; AveTimePerlter, represents the average time (in terms of ys) per iteration; and TotIterTimePPM, represents the total iteration time in terms of part per million of the overall encoding time. The test was run on a 2011 Laptop with a Quad-Core i17-2.70GHz CPU and 8GB memory. The measure results are summarized in Table V. It can be seen that on average, less than 10 times of Newton-Raphson iterations are required to perform QPC and the average time of each iteration is less than 1 ws. As the Qp increase, the AveNumlIterPerGOP slightly increases. Besides, the total time cost of these iterations is less than | part per mil- lion of the overall encoding time. Therefore, the computation overhead of Newton-Raphson method is negligible. proposed approach is negligible. Investigations of the Newton- Raphson method also show that, less than 10 iterations are required for most cases, which ensures a low computational complexity. Therefore, the coding time differences between the original HEVC and the proposed approaches are mostly determined by the changing of Qps, which may be positive or negative. Therefore, our scheme may even reduce the overall coding time in some cases.
Fig. 3. An example of an LD prediction structure with a GOP size of 4. COMPARISON OF DIFFERENT METHODS WITH HEVC LD
Fig. 3. An example of an LD prediction structure with a GOP size of 4. COMPARISON OF DIFFERENT METHODS WITH HEVC LD
COMPARISON OF DIFFERENT METHODS WITH HEVC LP Fig. 4. The proposed QPC method adapts to different @ values.
COMPARISON OF DIFFERENT METHODS WITH HEVC LP Fig. 4. The proposed QPC method adapts to different @ values.
Fig. 5. The proposed QPC method adapts to different SI and TI values. (a) & versus SI/O. (b) & versus TI/O.
Fig. 5. The proposed QPC method adapts to different SI and TI values. (a) & versus SI/O. (b) & versus TI/O.
Fig. 6. The proposed QPC method adapts in timeline.
Fig. 6. The proposed QPC method adapts in timeline.
Related topics:
Cognitive ScienceComputer ScienceMedicineElectrical and Electronic Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved