用於HEVC可調視訊編碼中估測模式相依之像素權重畫面內預測演算法

Abstract

A scalability extension to HEVC is currently under joint development of the ISO/IEC Moving Picture Experts Group and the ITU-T Video Coding Experts Group, offering two approaches, RefIdx and TextureRL, to performing inter-layer prediction. In the context of the TextureRL approach, this work proposes a mode-dependent pixel-based weighted intra prediction (MPWIP) for coding the enhancement layer (EL). The scheme first decomposes the EL predictor, constructed by the conventional directional prediction, and the collocated base-layer (BL) reconstructed block into their respective DC and AC components and then computes a weighted sum of both to form a better prediction signal using a pixel-based weighting scheme. The weighting factors to associate with different components are obtained by a least-squares fit to training data. It was observed that they depend strongly on the EL’s intra prediction mode and prediction block size, but are less dependent on QP settings under the common test conditions. The MPWIP offers an average BD-rate savings of 1.0% for the AI-2x configuration and 0.5% for AI-1.5x over the SHM-1.0 anchor, as compared to 0.3% and 0.1% to the max shown for each configuration respectively in the other prior works.

ISO/IEC運動圖像專家組和ITU-T的視頻編碼專家組目前正在聯合開發HEVC的可伸縮性延長，他們提供了RefIdx和TextureRL這兩種方法來執行層間預測。根據TextureRL的設計，我們提出了一種模式相關的基於像素的加權幀內預測（MPWIP）編碼的增強層（EL）的方法。此方案首先分解的EL預測，通過常規的預測，與並列配置的基本層（BL）構建重構到它們各自的直流分量和交流分量的塊，然後使用一個基於像素的加權方案來計算它們的加權總和，以達到更好的預測信號。用最小二乘法擬合對數據進行訓練，得到與不同的組件相關聯的加權因子。根據觀察，在常用的測試條件下，他們強烈依賴EL的幀內預測模式和預測塊的大小，但不依賴於QP設置。相較於SHM-1.0，MPWIP在AI-2倍的配置下節省了約1.0％的平均BD率，在AI-1.5的配置下節省了約0.5％的平均BD率。此外，跟其它現有工程中表現最優秀的比較起來，在AI-2倍的配置下節省了約0.3％的平均BD率，在AI-1.5的配置下節省了約0.1％的平均BD率。