可調式視訊編碼透過層際間(inter-layer)預測的技術，大幅改善編碼效能，但也大幅提高了計算複雜度。對於不同的視訊影片和編碼參數下，如何降低編碼的複雜度是增加可調式視訊編碼應用性的關鍵。本論文對於CGS可調性提出了一個以線上統計為基礎的快速模式決策演算法，包含為訓練階段和加速階段。在訓練階段，我們線上統計層際間的模式分佈、RD cost increase以及層際內RD cost increase關聯性以確保此統計能適應不同編碼設定下與不同視訊內容，並應用以此設計貝氏分類器和提早終止演算法以達成快速模式決策。在加速階段，依據貝氏分類器之loss functions可減少候選modes的數量並排序候選modes，提早終止演算法則進一步減少測試modes的數量。實驗顯示，我們提出的加速演算法可以節省73%左右的編碼時間，如果只考慮加速階段，則最多可節省78%左右的編碼時間，並且在不同的量化參數(QP)與影片內容時皆下幾乎不會造成PSNR的下降及位元率(bitrate)的上升。Duo to the high coding complexity of scalable video coding (SVC) that tests all candidate modes in RD optimization (RDO), in this thesis, we propose an on-line statistics based fast mode decision algorithm for SVC with CGS scalability. This algorithm consists of a training stage and an acceleration stage. At the training stage, the encoder gets the correlations of inter-layer mode distributions and intra and inter-layer RD cost increases. By this, we not only consider the probability distribution functions of optimal modes but also take the RD cost increase that caused by selecting a non-optimal mode into account. At the acceleration stage, the Bayesian classifier and an early termination algorithm are applied to reduce the candidate modes in RDO. By online statistics, the encoder can change the decision rule to reduce the RD performance loss for different sequences and QP sets. The experimental results show that our proposed algorithm can speed up to 73%. At the acceleration stage, our proposed algorithm can speed up to 78% with negligible RD performance loss.
