基於最佳停止理論所設計之可調視訊多層編碼控制的快速編碼模式決策演算法

Abstract

基於可調視訊編碼(SVC)之多層編碼控制(Multi-layer Encoder Control, MLEC)的架構，本論文提出一個快速編碼模式決策的演算法。此提出的演算法主要在於改善Zhao’s et al.應用最佳停止理論(Optimal Stopping Theory, OST)所提出的快速編碼模式決策演算法中所存在的問題。其演算法的問題是將所有編碼模式(mode)的效益都視為相同，只考慮編碼模式序列中下一個相對更好的編碼模式的位置和目前在序列中停止的位置相距多遠，但此距離並不能給予任何保證。不能保證這個相對更好的編碼模式可以得到多少RD-cost的下降，而且也不能保證需要花費多少運算複雜度。因此，本論文將以上兩點加以考慮，設計出一套更有效率的快速編碼模式決策演算法。 本論文將分析編碼模式配對序列中的不同停止點，若往下找到一組相對更好的編碼模式配對時能得到多少RD-cost下降量，以及分析不同的編碼模式配對在編碼時各需要多少運算複雜度。再將以上兩項資訊建為表格，提供在計算最佳停止點時所使用。除此之外，本論文還提出一個精進編碼模式決策的方法，利用統計與分析RD資訊找出較差的編碼模式配對，進一步執行精進程序。實驗結果顯示，本論文所提出的快速編碼模式決策演算法和徹底式搜尋(exhaustive search)所需編碼時間相比，節省了編碼時間逾80%~85%左右且有相當的RD效能。此外，本論文所提出的演算法和其他兩種不同的演算法相比，亦也有更好的穩定性，以及編碼品質。

This thesis proposed a fast mode decision algorithm for scalable video coding (SVC) multi-layer encoder control (MLEC). The proposed algorithm attempts to improve the problem of the fast algorithm proposed by Zhao’s et al. based on the optimal stopping theory (OST), in which the problem is the benefit obtained of the different mode is regarded as the same and only the distance considered between the stopping point and the position of the next relative best mode in the mode sequence. Unfortunately, this distance cannot guarantee the amount of RD-cost decrease in the next relative best mode; additionally, the amount of required computational complexity to the next relative best mode cannot be anticipated either. Thus, this thesis will take aforementioned points into consideration in order to create a more efficient fast mode decision algorithm. This thesis analyzed the stopping point in the mode pair sequence, which explore the amount of RD-cost decrease by finding the next relative best mode pair, and analyze the different amount of computational complexity between the different mode pairs. A table containing the above two points will also be provided for calculation of the optimal stopping point. Additionally, this thesis will proposed a more refined encoding method by utilizing statistics and data analysis to identify inferior encoding mode pairs and then refining them. Experimental result demonstrates that the proposed fast mode decision algorithm compares to the method of exhaustive search can achieve time saving about 80% to 85% with negligible quality degradation. Furthermore, the proposed method is superior to those of the other two types because of its stability and encoding quality.