於VVC視訊編碼畫面內針對編碼單元劃分模式之快速演算法

摘要

自2015年開始JVET (Joint Video Exploration Team)開討論起最新一代的視訊壓縮標準H.266/VVC討論最新的視訊壓縮標準H.266/VVC (Versatile Video Coding)。相較於前一代標準採用了QTMT (Quad Tree with nested Multi-type Tree coding block structure)的CU編碼結構。其支援最大128×128至最小4×4的方形以及矩形編碼區塊。該種結構能較好對視訊紋理做細分，提升編碼品質，但如此複雜的結構也將伴隨大量的演算法耗時，所以如何使用快速演算法使編碼品質和耗時達成平衡將是本論文的目標。

本論文提出基於特徵分析的QTMT快速演算法，該算法能分別減少二分樹劃分和三分樹劃分內可被利用的模式，其中二分樹與三分樹劃分的判斷結構皆相同，該演算法分為3部分，特徵圖建立與分析、傳統的分類方式與神經網路模型分的建立。首先，建立基於QTMT單位分區的特徵圖，並且利用該特徵圖生成編碼分區的特徵資料組。然後傳統的分析方法找出最佳的判斷式，將有著顯著資料的特徵組進行判斷。如果該特徵組有著細微變化，則特徵圖會進入提出的摺積神經網路模型進行分類。

關鍵字 : 多功能視訊編碼、編碼單位、快速演算法、畫面內編碼、特徵轉換、特徵分析、摺積神經網路模型

　

Fast CU Partitioning Algorithm for VVC Intra Coding

Abstract

Since 2015, JVET (Joint Video Exploration Team) has started to discuss the latest video compression standard H.266/VVC (Versatile Video Coding). Compared with the previous generation standard, the CU coding structure of QTMT (Quad Tree with nested Multi-type Tree coding block structure) is adopted. It supports square and rectangular coded blocks from a maximum of 128×128 to a minimum of 4×4. This structure can better subdivide the video texture and improve the coding quality, but such a complex structure will also be accompanied by a lot of time-consuming algorithms, so how to use fast algorithms to balance the coding quality and time-consuming will be the focus of this paper the goal.

This paper proposes a fast MT algorithm based on feature analysis, which can reduce the modes that can be used in BT partition and TT partition respectively. Among them, the judgment structure of binary tree and tri-tree partition is the same. There are three parts, the establishment and analysis of feature maps, the establishment of traditional classification methods and neural network models. First, a feature map based on the MT unit partition is established, and the feature map is used to generate a feature data set of the coding partition. Then, the traditional analysis method is used to find the best judgment formula, and judges the characteristic group with significant data. If the feature group has slight changes, the feature map will enter the proposed convolutional neural network model for classification.

Keyword: Versatile Video Coding (VVC), Coding Unit (CU), Fast Algorithm, Intra Coding, Spatial Feature, feature analysis, convolutional neural network model(CNN)