隨著科技進步影像的品質、解析度也隨之增加，資料量也越來越大，HEVC(High Efficiency Video Coding)又稱為H.265，採用更新的技術來降低位元率，連帶的也提升編碼計算複雜度。本論文採用深度學習與機器學習中的卷積神經網路CNN ( Convolutional Neural Network ) 和支持向量機SVM ( Support Vector Machine )，應用於HEVC編碼單元決策。將一個CTU分類成深度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HEVC遞迴運算處理編碼單元的方式，在特定深度提前終止後續編碼計算。另外在畫面內使用支持向量機的決策閥值，透過特定條件減少進入卷積神經網路的次數以利於節省編碼時間。整體平均BDBR上升至4.72%，編碼時間平均可節省75.22%，並探討在Random Access和Low Delay架構下性能比較，以及不同GOP大小產生的影響。;With the advancement of technology, the quality and resolution of images have also increased, and the amount of data becomes larger and larger. HEVC (High Efficiency Video Coding), also known as H.265, uses newer technology to reduce the bit rate, and the associated Improve coding computational complexity. This paper uses the convolutional neural network (CNN) and the support vector machine (SVM) in deep learning and machine learning that have flourished in recent years to apply it to HEVC coding unit decision-making. At the beginning of encoding, SVM is first used to classify the coding unit depth and prediction unit mode, and a CTU is classified into four categories: depth 0, depth 0~1, depth 0~2, and depth 0~3, and then the convolutional neural network is used. The network layer is subdivided downward. By using HEVC′s recursive operation to process coding units, subsequent coding calculations are terminated early at a specific depth. In addition, the decision threshold of the support vector machine is used to reduce the number of entries into the convolutional neural network through specific conditions to save coding time. The overall average BDBR increases to 4.72%, and the encoding time can be saved by 75.22% on average. Discuss the performance comparison between Random Access and Low Delay architectures, and the impact of different GOP size.
