Title Page
Abstract
Contents
I. Introduction 13
1.1. Motivations 13
1.2. Main contribution 18
1.3. Organization of Dissertation 21
II. Related Works 22
2.1. Backgrounds 22
2.1.1. Design of Integer DCT Transform Kernel 22
2.1.2. Quantization 28
2.1.3. Hierarchical Quadtree Coding Structure 32
2.2. Related Works 51
2.2.1. Review for Analysis of Transform Coefficients Distributions 51
2.2.2. Classical Model 55
2.2.3. Real Time Rate-Distortion Model 56
2.2.4. Distribution-based Rate-Distortion Model 57
2.2.5. No-Reference MSE Estimation 61
2.3. Quadtree Coding Structure of High Efficiency Video Coding 65
III. Statistical Analysis for HEVC Quadtree Coding 72
IV. Proposed Rate and Distortion Models 78
4.1. Estimation of Model Parameters 81
4.2. Proposed Distortion Model 84
4.3. Proposed Rate Model 91
V. Application of Distortion Models to No-Reference PSNR Estimation for Quadtree Coding of HEVC 95
5.1. Backgrounds 95
5.2. Proposed MSE (PSNR) Estimation 96
5.2.1. Laplacian parameter estimation from the quantized coefficients 98
5.2.2. MSE estimation for intra-coded block 99
5.2.3. Parameter estimation for all-zero coefficient blocks 103
5.2.4. No-reference PSNR estimation algorithm 105
VI. Experimental Results 107
6.1. Experimental Results for Rate and Distortion Models 107
6.2. Experimental Results for No-reference PSNR Estimation 131
VII. Conclusions and Future Researches 136
References 140
국문요약 147
Table II-1. Approximation errors of the order-16 ICT kernels and transform coding gains (ρ=0.85) 25
Table II-2. Comparisons of order-16 ICT kernels in terms of approximation errors 27
Table II-3. Comparisons of order-16 ICT kernels in terms of coding gains 27
Table II-4. Transform types of THE HVBT for MB modes 37
Table II-5. Low complexity RD based variable block transform 41
Table II-6. Performance comparisons between HVBT and ST in terms of BDBR, BDPSNR 42
Table II-7. Comparisons between conventional codec and emerging HEVC 66
Table III-1. Proportions of CU block in various depth levels and for different coding types 76
Table III-2. Average variances of transform coefficients for different depth levels of the CU and coding types of CU blocks 77
Table IV-1. Bit shift numbers in H.264/AVC and HEVC 81
Table V-1. Proposed no-reference PSNR estimation algorithm 105
Table VI-1. Performance of distortion estimations: Mmse is the mean for actual MSEs during encoding.(이미지참조) 108
Table VI-2. Performance of rate estimations: Mbits is the mean for actual number of bits during encoding.(이미지참조) 109
Table VI-3. Performance of distortion estimations when LCU size is 32×32: Mmse is the mean for actual MSEs during encoding.(이미지참조) 130
Table VI-4. Performance of rate estimations when LCU size is 32×32: Mbits is the mean for actual number of bits during encoding.(이미지참조) 130
Table VI-5. Performance of distortion estimations when LCU size is 16×16: Mmse is the mean for actual MSEs during encoding.(이미지참조) 131
Table VI-6. Performance of rate estimations when LCU size is 16×16: Mbits is the mean for actual number of bits during encoding.(이미지참조) 131
Table VI-7. Comparison for no-reference PSNR estimation 134
Figure I-1. Rate and distortion model to be resided in a hybrid video codec 13
Figure I-2. Architecture of No-reference PSNR estimation 14
Figure I-3. Applications of the rate and distortion models 14
Figure I-4. An example of CU and TU partitions in the hierarchical quadtree structure of HEVC 16
Figure II-1. Approximation comparisons of different order-16 ICT kernels for DCT 28
Figure II-2. DZ+UTQ for various rounding offset parameters 29
Figure II-3. Quantization parameter (QP) and quantization step size in HEVC 30
Figure II-4. Transform structure in H.264/AVC 35
Figure II-5. An illustration of the HVBT 35
Figure II-6. Modeling of pixel correlations for 2-D Input signal with AR (1) process 39
Figure II-7. Transform coding gains for the HVBT scheme, and two single-type transforms for (pv¹,pn²) in block 1 and block 2 in Figure II-6.(이미지참조) 40
Figure II-8. RD curves for various test sequences 45
Figure II-9. RD performance comparisons for the proposed kernel and others 47
Figure II-10. Proportion of different transform block types for various spatial resolutions 48
Figure II-11. Comparison of selected transform types between H.264/AVC and the HVBT with a cropped part of the 5-th frame of BasketballDrill (832×480) for QP=20 and QP=28. 50
Figure II-12. Coding Unit (CU), Prediction Unit (PU) and Transform Unit (TU) in HEVC 68
Figure II-13. Quadtree partitioning and side information of CU, PU and TU 69
Figure II-14. Quadtree partitioning and side information of CU, PU and TU 70
Figure II-15. Quadtree partitioning and side information of CU, PU and TU 71
Figure III-1. Quadtree partitions of CU and TU: (a) CU partitions for RaceHorses (416×240) - Blue lines: LCU, Green lines: CU blocks; (b) Quadtree partitions of CU and TU block in a LCU, Red lines: TU blocks, Yellow lines: CU blocks 73
Figure III-2. Statistical analysis of pixel data for CU levels and coding types for predicted residual 75
Figure IV-1. PDF comparisons for fCUk(l) and a single PDF(이미지참조) 83
Figure IV-2. Linear relations between actual distortions and ratios of relative block size. 87
Figure IV-3. Plots of variance fluctuations, relative ratios and predicted MSEs for inter- and intra-coded CU blocks by the estimated PDF based MSE and the proposed distortion model in (39) 90
Figure IV-4. Average qλk values for different QPs and CU depth levels. (a) 832×480, (b) 1280×720 and (c) 1920×1080(이미지참조) 94
Figure IV-5. Relation between actual bits and entropy in (4.29) for CUk(이미지참조) 94
Figure V-1. Variance variations for CUk during encoding(이미지참조) 101
Figure V-2. Average variances of transform coefficients for CU depth levels 101
Figure V-3. β versus QPs 102
Figure V-4. Variances versus CU depth levels. (a) an illustration of model parameter estimation for all zero coefficient CU blocks using exponential regression; (b) variance curve of actual coefficient values and its exponential regression. 104
Figure VI-1. BlowingBubbles with complex and various texture characteristics and Vidyo1 with homogeneous and simple texture characteristics.(이미지참조) 111
Figure VI-2. Rate and distortion prediction performances vs. variances and proportions of CUk for BlowingBubbles sequence encoded with QP=28(이미지참조) 113
Figure VI-3. Rate and distortion prediction performances vs. the variances and proportions of CUk for Vidyo1 sequence encoded with QP=28(이미지참조) 115
Figure VI-4. Performance comparisons for distortion estimation 117
Figure VI-5. Performance comparisons for distortion estimation (BQMall_ 832×480) 119
Figure VI-6. Performance comparisons for distortion estimation (RaceHorses_832×480) 120
Figure VI-7. Performance comparisons for distortion estimation 122
Figure VI-8. Performance comparisons for rate estimation 124
Figure VI-9. Performance comparisons for rate estimation 125
Figure VI-10. Performance comparisons for rate estimation 127
Figure VI-11. Performance comparisons for rate estimation 128
Figure VI-12. Actual PSNR vs. No-reference PSNR estimation 133
Figure VI-13. Actual PSNR vs. No-reference PSNR estimation 135