在這網路快速進步的時代，對於高解析度影像的需求不斷提升，高解析度代表著資料量相對龐大，HEVC/H.265採用編碼單元(Coding Unit,CU)、預測單元(Prediction Unit,PU)、碼率失真最佳化(Rate-Distortion Optimization)等等，這些先進的編碼技術提高了壓縮率，但運算複雜度卻也大幅的增加，本論文結合卷積神經網路與支持向量機應用於編碼單元深度決策。首先在編碼一開始使用支持向量機將編碼單元分類為只做深度0、深度0~1、深度0~2、深度0~3四種類別，再各別使用卷積神經網路依據在支持向量機已取得的畫面間預測移動向量值做為特徵(Feature)，判斷是否需要提前終止，提前終止的區塊只會進行一次深度的編碼，且因為移動向量值為特徵複用，進而節省編碼所需花費的運算時間。在只進行64x64編碼決策的情況下，實驗結果與HEVC進行比較，平均BDBR上升1.32%的情況下，編碼時間節省46.84%。;In the era of rapid Internet advancement, the demand for high-resolution images continues to increase. The use of high-resolution images implies that a large amount of data is resulted. HEVC/H.265 adopts advanced encoding techniques such as Coding Unit (CU), Prediction Unit (PU), and Rate-Distortion Optimization to improve the compression ratio of data; however, such approach also increases the computational complexity significantly. In this thesis, Convolutional Neural Network (CNN) was combined with Support Vector Machine (SVM) and applied to the depth decision of coding unit. At the beginning of the coding process, Support Vector Machine was used to sort the coding units into four categories of depth 0, depth 0~1, depth 0~2 and depth 0~3. Convolutional Neural Network was then used to determine whether early termination is needed based on the inter prediction motion vector value obtained by the Support Vector Machine as a feature. The block that terminates early will only be deep-coded once. Since the motion vector value is feature multiplex, it reduces the computation time required for coding. For 64x64 coding decision, the experimental results were compared with HEVC, showing that the coding time was reduced by 46.84% when the average BDBR was increased by 1.32%.
