用於HEVC編碼單元之快速決策演算法 與 結合移動向量與DCT之H.264編碼器優化

Abstract

由於高解析度影像應用的需求，視訊編碼在3C產品中是不可或缺的技術，例如行動電話、高畫質電視、藍光光碟機。進階視訊編碼(Advanced Video Coding, AVC/H.264)是目前商業產品中，廣泛採用的壓縮標準格式。為了達到更高的編碼效率，國際組織JCT-VC正在進行下一代標準的制定，即高效率視訊編碼(High Efficiency Video Coding, HEVC) 。 相較於進階視訊編碼，雖然高效率視訊編碼的複雜度提升許多，但是在相似的影像品質下，可以增加近一倍的壓縮效率。 此論文包含兩個研究主題：第一個主題是改善進階視訊編碼中，整數精確度的移動估測以增進編碼效能；第二個主題是關於高效率視訊編碼的編碼單元(Coding Unit, CU)大小的快速決策，以達到降低編碼器複雜度的目標。在進階視訊編碼中，整數移動估測的失真項，是以區塊之絕對誤差總和(The Sum of the Absolute Distortion, SAD)來計算，但是此方法並不能完全反應最後結果的失真。為了在相似的畫面品質下，進一步節省位元率，我們提出迭代的位元率-失真(Rate-Distortion, R-D)計算方式，以選擇較佳的移動向量。我們將此演算法實現於JM18.0，用許多組MPEG測試影像來檢驗此方法的效能，並將執行結果和原始JM做法的結果進行比較。雖然JM18.0是發展已久的優化編碼器，我們仍可從中節省1.1%至4.2%的位元率，但代價是增加45%的運算複雜度。 另一方面，高效率視訊編碼在傳統的編碼流程中，增加了編碼單元四元分割樹的構造。彈性的編碼單元設計提升了編碼效率，但相較於進階視訊編碼傳統的巨區塊(Macroblock, MB)結構而言，編碼複雜度提升不少。因此我們設計快速演算法以有效率地建造出編碼單元四元分割樹，其中演算法包括分裂決策、終止決策。這些快速編碼單元大小決策參考週遭相關的編碼單元之切割資訊以進行判斷。此外，我們設計額外的工具以增進我們提出的演算法效能，其中包含畫面層級加速控制和跳過決策後的快速預測單元判斷。最後，我們分析提出的快速演算法，並和HM5.0中的兩種快速演算法進行比較，以找出有效率的結合方法。相較於HM5.0的原始設定，我們提出的快速演算法，經過多組高解析度的影像測試，可以節省高達49%的整體編碼時間，但平均損失0.06dB的峰值信噪比(PSNR)。

With the growing demand for high resolution video applications, video coding is an indispensable element in many 3C products, such as mobile phone, DTV, and BD player. Today, Advanced Video coding (AVC/H.264) is one of the most popular video formats in commercial applications. Aiming at higher compression efficiency, the international JCT-VC is currently developing the next generation standard, High Efficiency Video Coding (HEVC). With a much higher encoder complexity, HEVC is able to achieve a 50% bitrate reduction compared to H.264/AVC. This thesis has two topics, one is the enhanced motion estimation (ME) for AVC/H.264 and the other is the fast coding unit (CU) decision for HEVC. In H.264, the sum of the absolute difference (SAD) is used as the distortion term in ME, but it does not reflect the final coding distortion. To achieve further bitrate reduction, we propose an enhanced motion vector selection method based on the iterative R-D calculation. We compare the proposed method with the original H.264/AVC JM18.0 reference software on several MPEG test sequences. Although JM18.0 is a highly optimized scheme, we can still obtain a BD-rate improvement from 1.1% to 4.2% but with additional 45% complexity increase. In HEVC, the CU quadtree structure is added to the traditional fixed size macroblock. With flexible CU size selection, the coding efficiency increases but the complexity of HEVC becomes much higher than that of AVC/H.264 fixed macroblock (MB) structure. To reduce computational complexity, we propose a fast algorithm, which includes the splitting decision and the termination decision, in building the CU quadtree. The fast CU size decision of the current CU makes use of the size information of its neighboring CUs. Furthermore, we design the additional tools to enhance the performance of the proposed algorithm. The additional tools include the frame level acceleration and the fast PU size decision after the splitting decision. At the end, we compare it with the existing fast algorithms in HM5.0 and find an efficient way to blend them together. In comparison to the original HM5.0, our method saves the overall encoding time up to 49% with 0.06 dB average PSNR drop.