遍佈即時多媒體系統與技術---子計畫二：可調視訊編碼之適應性動態精細改進和編碼

Abstract

在本子計劃，我們提出一個適應性動態精細改進演算法用以改進目前MPEG-4 Part 10 Amd.1 可調視訊編碼標準中的細緻可調性。 目前的可調視訊編碼標準提供兩種型式的畫質可調性。其中一個是粗略可調視訊 編碼，另一個則是細緻可調視訊編碼。粗略可調視訊將視訊壓縮成多個畫質可調層並 藉此提供一組有限的畫質可調範圍。細緻可調視訊編碼則進一步的在每一個畫質可調 層利用內嵌式的Entropy 編碼產生一組近似無限多個的畫質可調範圍。除了使用不同的 量化設定，現有的兩種可調視訊編碼有以下幾點主要不同： .. 粗略可調視訊編碼允許各別的畫質可調層擁有獨自的動態向量，相反的細緻 可調視訊編碼則在不同的畫質可調層共用相同的動態向量。 .. 粗略可調視訊編碼允許適應性的畫質可調層間估測，相反的細緻可調視訊編 碼則固定強制使用畫質可調層間估測。 .. 粗略可調視訊編碼利用在MPEG-4 精進可調視訊編碼中的CABAC 編碼畫質 可調層，相反的細緻可調視訊編碼則使用一個可調的循環方塊編碼。 當可調範圍變大時，我們發現細緻可調視訊編碼會有較差的編碼效能。其原因如 下： .. 固定的動態向量並不是對所有位元率都是最佳。 .. 固定強制使用畫質可調層間估測並不是對所有位元率都是最佳。 基於以上原因，我們提出一個適應性動態精細演算法用於可調視訊編碼標準中的細 緻可調性。具體來說，當可調範圍變大時，我們允許在畫質可調層間有多個動態向量。 部份可由基底層產生，其他則可藉由畫質可調層產生。每一個畫質可調層可以有各自的 動態向量，而多個畫質可調層也可以共享相同的動態向量。 適應性動態精細演算法所帶來的編碼效益會付出額外的位元率用於動態向量估 測殘餘值和畫質可調層間估測模式。在細緻可調性中，額外所需的位元率將更加的顯 著，因此會限制最後的編碼效益。所以，在本子計劃，我們想探討是否對不同類型的可 調性應該有不同的編碼方式用在動態向量估測殘餘值和畫質可調層間估測模式。同 時，是否存在著較有效率的前後文參照模型。除此以外，我們也想知道是否可以用單一 的Entropy 編碼套在不同類型的可調性。 本子計劃的年度目標包括以下幾點: (1) 提出適應性動態精細演算法用於細緻可調性 (2) 開發一個統一的Entropy 編碼套在不同類型的可調性 (3) 利用可調視訊編碼軟體評估所提的方法 (4) 利用所提的方法參與可調視訊編碼標準制定

In this project, we propose an adaptive motion field refinement scheme to improve the fine granularity scalability in MPEG-4 Part 10 Amd. 1 Scalable Video Coding. In SVC, there are two types of SNR scalability which are coarse granularity scalability (CGS) and fine granularity scalability (FGS). The CGS provides a finite number of quality levels by compressing the video into multiple quality layers. The FGS further introduces an embedded entropy coding scheme in each quality layer to produce an infinite number of quality levels. Except for the step size between successive quantization, the major differences of these two schemes include the following: .. CGS allows each quality layer has its own motion field while FGS uses a fixed motion field for all quality layers. .. CGS employs an adaptive inter-layer prediction while FGS always enforces inter-layer prediction. .. CGS uses the CABAC in MPEG-4 AVC as entropy coder while FGS employs a scalable cyclical block coding. As the bit rate range of SNR scalability increases, we find that FGS suffers from poor coding efficiency because of the following reasons: .. A fixed motion field is not optimal for all bit rates. .. An enforced inter-layer prediction is not optimal for all bit rates. Thus, in this project, we introduce an adaptive motion refinement scheme for FGS in SVC. Specifically, there are more than one motion vector fields (MVFs) when the bit rate range is wide. One MVF is generated at base layer and others are generated at enhancement layers. Each layer can have its own MVF and several layers can also share one MVF. The coding gain of adaptive motion filed refinement comes with the cost of extra bits on the prediction residue of MVFs and the inter-layer prediction mode. In the FGS, these extra bits are transmitted more frequently, which limits the coding gain from adaptive motion refinement. Thus, we would like to investigate if there should be different representation of the inter-layer prediction mode for CGS or FGS. Also, if there exists more efficient context models for the prediction mode as well as the residue of motion vector. We would like to evaluate how much gain can be achieved by having different representations. In addition, we also want to know if there exists a unified entropy coding scheme that can be applied in both CGS and FGS scenarios. The yearly goals of this project include the following: (1) Propose an adaptive motion refinement scheme for FGS. (2) Develop a unified entropy coder that can be used in both CGS and FGS. (3) Evaluate the proposed schemes based on the latest joint scalable video model. (4) Joint the standardization activities of SVC standard with the proposed schemes.