基於支持向量機之HEVC畫面內編碼單位快速決策演算法
由JCT-VC (ISO/IEC MPEG和 ITU-TVCEG)所制定的最新一代視訊壓縮標準High Efficiency Video Coding (HEVC)，其編碼效率相較於目前主流H.264視訊壓縮標準有顯著提升。延續H.264的巨區塊架構(Macroblock)，HEVC將基本編碼區塊改為編碼單元(Coding unit, CU)，並採用樹狀編碼結構(Quad-tree)提供更多編碼區塊大小以適應畫面特性，但此種樹狀架構也大幅增加了計算複雜度；而從視訊解析度不斷提升的演進來看，相較於畫面間編碼(Inter coding)，畫面內編碼(Intra coding)更能針對畫面中高移動量的部份以較精準的方向模式(Intra mode)去預測，因此發展畫面內CU深度決策快速演算法有其必要性。
本論文提出一個應用於畫面內編碼的CU深度快速決策演算法，擷取四種空間上的相關性以及原始畫面的資訊為特徵(Feature)，包含鄰近CU深度、邊界像素差值、像素變異數以及邊緣點數量，利用類神經網路分析這些特徵與CU切割與否的關聯性，接著使用支持向量機(Support vector machine, SVM)以機器學習的方式，針對不同CU深度歸納出一套快速CU深度決策演算法，減少位元-失真最佳化程序(Rate-Distortion
Optimization)所帶來的龐大運算量。實驗結果顯示，在些微增加位元率的情況下，利用本演算法平均可以減少46.5%，最高至58.9%的總編碼時間。
SVM based fast intra CU depth
decision for HEVC
The latest video coding
standard, High Efficiency Video Coding (HEVC) adopted quad-tree based coding
unit (CU) as an extension of macroblock (MB) in H.264/AVC, had achieved much
higher coding efficiency. However, the significant increase of complexity due
to the advanced encoding structure can’t be neglected.
In this paper, an SVM based
fast CU depth decision is proposed to reduce the computational complexity. It
is convenient to develop the criterion of early CU splitting and termination by
applying SVM. Neural network is used to analyze the features extracted from
spatial domain and pixel domain, including neighboring CU depth, boundary pixel
difference, pixel variance and number of edge points, then proper features are
selected as the input of SVM.
In addition, intra coding
is more efficient than inter coding when video resolution becomes higher since
it is hard to perform motion estimation well in a limited area when strong
motion exists. Therefore, we aimed to develop a SVM based fast intra CU depth
decision for all-intra coding structure. The experiment results show that this
fast algorithm provides 58.9% encoding time saving at most, and 46.5% encoding
time saving on average compared to HM 12.1.
Index terms – HEVC, All intra, CU, fast algorithm, SVM