| 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，VVC採用了一些更新穎的技術，如矩形編碼樹單元、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，即卷積神經網路與隨機森林分類器，將其應用於VVC編碼單元編碼區外的劃分。不同於原始VVC遞迴運算編碼單元碼率失真成本，本論文在編碼一開始時先使用支持向量機及卷積神經網路，將方形編碼單元區塊做出劃分，再利用隨機森林分類器向下細分矩形編碼單元區塊，分類完成的區塊將只會進行一次的編碼，藉此大幅節省編碼所需時間，後續再透過隨機森林決策輔助原始VVC篩選預測模式的方式，將整體計算縮減至不到兩成。後續在解碼端則引入三通道殘差神經網路架構，以不同的資訊去補償我們在編碼端的失真。以此實現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。實驗結果與VVC相比，整體平均BDBR下降1.63%的情況下，整體編解碼時間大約可以節省51.48%。;In this ever-changing era, with the advancement of the Internet and the development of technology, people will never stop pursuing higher-quality things, and the same is true for high-resolution images. In order to compress these huge videos more efficiently data volume, VVC adopts some more novel technologies, such as rectangular coding tree unit, rate-distortion optimization, etc., but at the same time, it also causes an increase in the complexity of coding calculations. This paper combines the very popular in recent years Deep learning and machine learning, namely convolutional neural networks and random forest classifiers, are applied to VVC coding unit depth decisions. Different from the original VVC recursive operation coding unit rate distortion cost, this paper first uses support vector machine and convolutional neural network to divide the square coding unit blocks at the beginning of coding, and then uses random forest classifier to Subdividing the rectangular coding unit block, the classified block will only be coded once, thereby greatly saving the time required for coding, and then using random forest decision-making to assist the original VVC to filter the prediction mode, reducing the overall calculation to Less than 20%. Subsequently, a three-channel residual neural network architecture is introduced at the decoding end to compensate our distortion at the encoding end with different information. In this way, the concept of distributed video coding is realized, and the fast prediction mode is combined with post-processing at the decoding end to compensate for image quality. Experimental results Compared with VVC, when the overall average BDBR is reduced by 1.63%, the overall side decoding time can be saved by about 51.48%. |
