1. A Frame-Level Rate Control Scheme Based on Texture and Nontexture Rate Models for HEVC IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 24, NO. 3, MARCH 2014 BUMSHIK LEE, MUNCHURL KIM, TRUONG Q. NGUYEN 1

2. Overview Introduction Texture and Nontexture Rate Models For HEVC Proposed Frame-level Rate Control For HEVC Experimental Results Conclusion 2

3. Introduction  In the current HEVC, more flexible hierarchical-block structures are adopted with quadtree partitions and higher depth level, which are composed of CU, PU and TU. 3

4. Introduction(cont.)  Various block sizes of prediction and transform in hierarchical structures often result in very different statistical characteristics of the residual signals.  So it is difficult to capture various statistical characteristics of residual signals with only one single rate-quantization model.  But there have been seldom works on the rate control for these structure like HEVC codecs.  Furthermore, in HEVC, the relative portion of texture bits has been reduced compared to that of nontexture bits. Therefore it becomes important to separately model the texture and nontexture rates for rate control. 4

5. Overview Introduction Texture and Nontexture Rate Models For HEVC Proposed Frame-level Rate Control For HEVC Experimental Results Conclusion 5

6. Why Texture and Nontexture Model Fig.2 Quadtree partitions of CU and TU. Optimal depth can be determined based on a RDO. 6

7. Why Texture and Nontexture Model(cont.) Average variances of transform coefficients for different depth levels of the CU and coding types Table I 7

8. Rate Modeling for Texture Bits 8

9. Rate Modeling for Texture Bits(cont.)  The histograms of the DCT coefficient values accept the chi-square test with a confidence interval 95%. By test above, we can justify that the DCT coefficient values well follow the Laplacian PDFs.  Thus, we model the source distribution of the transform residues for CU as 9

10. Rate Modeling for Texture Bits(cont.)  As shown in table I, a rate model in a frame level is proposed by taking into account the resulting bit amounts from the CU blocks of different CU categories. 10

11. Rate Modeling for Texture Bits(cont.) 11

12. Rate Modeling for Texture Bits(cont.)  The entropy can be expressed in closed form by substituting (5) and (6) into (4) 12

13. Rate Modeling for Texture Bits(cont.)  Finally, by substituting (9) into (3), the proposed rate model for texture can be rewritten as The sum of the block sizes(pixels) of nonSKIP mode in each category 3 13

14. Rate Modeling for Texture Bits(cont.) 14

15. Analysis and Rate Modeling for Nontexture Bits  Nontexture data, which includes motion vectors, quadtree split information for TU and CU, signaling information of prediction models, filter signaling information etc., becomes more significantly important due to the adoption of various coding tools. 15

16. Analysis and Rate Modeling for Nontexture Bits(cont.)  Most of the conventional rate control schemes employ a nontexture data collected from a few previous frames [7],[9],[13],[17],[20],[21].  In [14], the nontexture bits are predicted by using the number of nonzero motion vectors.  In [15], a unified linear rate-quantization model (R-Q model) is used to estimate the texture and non-texture rates.  But unpredictable nature of nontexture bits may influence the entire rate control performance, authors thoroughly analyze the characteristics of nontexture data in various CU depths. 16

17. Analysis and Rate Modeling for Nontexture Bits(cont.) Nontexture bits 17

18. Analysis and Rate Modeling for Nontexture Bits(cont.)  Since the CU in the deepest coding depth level CU 3 has the smallest partitioned blocks, many MV occurs.  The dominant portions of nontexture data types are different depend on CU categories. 18

19. Analysis and Rate Modeling for Nontexture Bits(cont.) 19

20. Analysis and Rate Modeling for Nontexture Bits(cont.) 20

21. Overview Introduction Texture and Nontexture Rate Models For HEVC Proposed Frame-level Rate Control For HEVC Experimental Results Conclusion 21

22. Proposed Frame-level Rate Control for HEVC  In this section, a rate control scheme is proposed based on the proposed rate models in above sections. For a GOP level bit allocation is 22

23. Proposed Frame-level Rate Control for HEVC 23

24. Proposed Frame-level Rate Control for HEVC(cont.) 24

25. Proposed Frame-level Rate Control for HEVC(cont.)  Issue of determining an initial QP value in each GOP : 25

26. Proposed Frame-level Rate Control for HEVC(cont.) 26

27. Overview Introduction Texture and Nontexture Rate Models For HEVC Proposed Frame-level Rate Control For HEVC Experimental Results Conclusion 27

28. Experimental Results  Perform experiments under environment of Intel Core i-7 CPU @3.40GHz with 8.0GB memory, and 64-bit Windows 7 operating system.  11 Test sequences  Comparison Model : HM10.0 rate control[30], Chen’s method[13] and JCTVC-H2013[24]  Comparison with PSNR, PSNR standard deviation, accuracy of target bitrates, visual quality, and buffer status level. 28

29. 29

30. Experimental Results(cont.) (a) BlowingBubbles (416x240) (b) PartyScene (832x480) 30

31. Experimental Results(cont.) (c) RaceHorses(832x480)(d) BQTerrace(1920x1080) 31

32. Experimental Results(cont.) 32

33. Experimental Results(cont.) (a) BlowingBubbles (416x240) (b) PartyScene (832x480) 33

34. Experimental Results(cont.) (c) RaceHorses(832x480)(d) BasketballDrive(1920x1080) 34

35. Experimental Results(cont.) 35 Fig. 11. PSNR and visual quality comparisons during encoding. (PartyScene 832x480) (e) Proposed rate control (28.88 dB) (f) HM10.0 rate control (28.55 dB) (g) Chen’s method [13] (26.54 dB) (h) JCTVC-H-213 [24] (26.99 dB)

36. Experimental Results(cont.) 36 Fig. 12. Coded number of bits for each frame for PartyScene sequence. (a) Frame by frame buffer levels for PartyScene sequence encoded at 1Mb/s. (b) Number of coding bits during encoding.

37. Overview Introduction Texture and Nontexture Rate Models For HEVC Proposed Frame-level Rate Control For HEVC Experimental Results Conclusion 37

38. Conclusion  In this paper, a frame-level rate control scheme is proposed based on novel rate models for texture and nontexture data for HEVC.  The proposed texture rate models takes the statistical characteristics of transform coefficient residues into account by using multiple Laplaican PDFs for different CU categories in various depth.  The proposed nontexture rate models can precisely estimate nontexture bits based on the linear relation between total nontexture data and the dominant nontexture data.  The experimental results show that the proposed method have higher PSNR and a lower PSNR standard deviation. 38