| 摘要: | 在這日新月異的時代，隨著網路的進步以及科技的發達，人們對於追求更高品質的事物始終不會停滯，對於高解析度的影像也是如此，為了能夠更有效率的壓縮這些巨大的視訊資料量，HEVC採用了一些更新穎的技術，如編碼樹單元、碼率失真最佳化等等，但於此同時也造成了編碼計算複雜度的提升，本論文結合近幾年來十分熱門的深度學習與機器學習，即卷積神經網路與支持向量機，將其應用於HEVC編碼單元深度決策。不同於原始HEVC遞迴運算編碼單元深度0至3，本論文在編碼一開始時先使用支持向量機將編碼單元分成簡單區塊與複雜區塊，再利用卷積神經網路分層向下細分，分類完成的區塊將只會進行一次深度的編碼，藉此大幅節省編碼所需時間。而後進一步將支持向量機的決策值，透過額外資訊減少進入卷積神經網路的次數便提前完成分區，實驗結果與HEVC相比，整體平均BDBR上升1.5%的情況下，編碼時間大約可以節省64%，後續再導入分散式視訊編碼的概念，結合快速預測模式與解碼端之後處理補償影像品質。;In this ever-changing era, with the advancement of the Internet and the development of technology, people will never stop pursuing higher-quality things, as well as high-resolution images. In order to be able to compress these huge videos more efficiently The amount of data, HEVC uses some newer technologies, such as coding tree units, rate distortion optimization, etc., but at the same time it also causes the increase in the complexity of coding calculations. This paper combines deep learning and machine learning, which have been very popular in recent years, that is, convolutional neural networks and support vector machines, are applied to HEVC coding unit depth decision. Unlike the original HEVC recursive operation coding unit depth 0 to 3, at the beginning of this paper, the support vector machine is used to divide the coding unit into simple blocks and complex blocks, and then the convolutional neural network is used to layer down , The classified blocks will only be coded once in depth, thereby greatly saving coding time. Then, the decision value of the support vector machine is further used to reduce the number of entering the convolutional neural network through additional information to complete the partition in advance,compared with HEVC, the overall average BDBR is increased by 1.5%, and the encoding time can be saved by about 64%.Finally, introduce the concept of decentralized video coding, combined with fast mode prediction and post-processing to compensate the image quality. |
