隨著科技日新月異及人們嚮往高解析度所帶來的影像品質，因此高解析度的影像產品也跟著與日俱增，以至於為了能夠有效壓縮高解析度影像膨大資料量，HEVC(High Efficiency Video Coding)採用了許多更新穎的技術來降低位元率，例如:畫面內預測、畫面間預測、碼率失真最佳化等等，但也同時造成了編碼計算複雜度提升。而本論文利用近幾年蓬勃發展起來的深度學習與機器學習中的卷積神經網路 CNN ( Convolutional Neural Network ) 及 支持向量機 SVM ( Support Vector Machine ),將其應用於HEVC編碼單元階段的決策。本論文在編碼一開始時先使用SVM來對編碼單元深度以及預測單元模式做分類，編碼單元以畫面間預測的移動向量值的資訊、合併模式的CBF、鄰近區塊深度資訊作為特徵(Feature)將一個CTU分類成只處理深度0、深度0~1、深度0~2、深度0~3四種類別，再利用卷積神經網路分層向下細分。藉由原本HEVC遞迴運算處理編碼單元的方式，在特定深度的編碼提前終止後續的編碼計算，以此節省後續深度所需計算時間達成編碼端縮減時間。最終實驗結果顯示，與HEVC相比，整體平均BDBR上升0.89%的情況下，編碼時間大約可以節省44%。;With the rapid development of technology and the desire for high resolution, the image products with high resolution are increasing. In order to effectively compress the volume of high-resolution image expansion data, HEVC (high efficiency video coding) adopts many more novel technologies to reduce bit rate, such as: in-picture prediction, inter picture prediction, and so on The optimization of bit rate distortion and so on also leads to the increase of the complexity of coding calculation. In this paper, we use convolutional neural network (CNN) and support vector machine (SVM) in deep learning and machine learning to make decision in hevc coding unit stage. At the beginning of coding, this paper uses SVM to classify coding unit depth and prediction unit mode. Coding unit classifies a CTU into four categories: depth 0, depth 0 ~ 1, depth 0 ~ 2 and depth 0 ~ 3 based on the information of motion vector value predicted between pictures, CBF of merging mode and depth information of adjacent blocks, Then the convolution neural network is used to subdivide downward. By using the original hevc recursive operation to process the coding unit, the subsequent coding calculation is terminated in advance at a specific depth of coding, so as to save the calculation time of the subsequent depth and reduce the coding time. The final experimental results show that compared with HEVC, when the overall average bdbr increases by 1.07%, the encoding time can be saved by about 45%.
