以深度學習改善HEVC插值器品質;CNN-based HEVC interpolation filters

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/83839

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/83839

Title:	以深度學習改善HEVC插值器品質;CNN-based HEVC interpolation filters
Authors:	黃琮閔;Huang, Cong-Ming
Contributors:	通訊工程學系
Keywords:	高效率視訊編碼;支持向量機;深度學習;畫面間預測;運動估計;HEVC;Support Vector Machine;Deep Learning;Inter Prediction;Motion Estimation
Date:	2020-07-31
Issue Date:	2020-09-02 17:12:23 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著人們要求視覺上的享受，網路和多媒體技術不斷進步，2013國際標準組織制訂了新的視訊壓縮標準，HEVC/H.265(High Efficient Video Coding)，HEVC視訊壓縮率可以提高到H.264兩倍以上，且畫質更優於H.264，相對的HEVC計算複雜度也提高了很多，原本H.264的插值濾波器，在亮度部分二分之一使用6-tap的濾波器、四分之一使用雙線性插值濾波器，而HEVC的插值濾波器，在亮度部分二分之一使用8-tap的濾波器、四分之一使用7-tap的濾波器，在插值濾波器的運算時間HEVC就比H.264多了兩倍的時間，而HEVC採用固定插值濾波器是根據信號處理理論設計的，前提假設視頻信號是理想的低通信號。但是視頻信號不一定是低通的並且不是固定的，近年來，基於深度學習的方法已被廣泛使用，並在圖像和視頻處理中獲得了顯著的效果，而本篇論文一開始先使用支持向量機，先將訓練資料分為四個次群，分別為四個次群訓練模型，並且根據消息理論，越多的側面消息，是能有效幫助到訓練的，本篇論文使用了SVM Features Mask與殘差做為我們的側面消息，並形成我們的雙輸入模型與三輸入模型，以三輸入模型來說，與HEVC相比BDBR可以下降1.33%，編碼時間節省了13.39%，但是編碼能節省時間是因為Liu[18]的支持向量機編碼單元快速決策演算法，實際上卷積神經網路是增加編碼時間的，所以我們提出了一個想法，使用最簡單的的雙線性插值濾波器，並且最後使用我們提出的架構做一個改善，以三通架構來說，BDBR可以下降約1.09%，編碼時間能節省20.64%。;As people demand visual enjoyment, network and multimedia technologies continue to progress, 2013 International Standards Organization has formulated a new video compression standard, HEVC/H.265 (High Efficient Video Coding), HEVC video compression rate can be increased to H. 264 is more than twice, and the image quality is better than H.264. The relative HEVC calculation complexity is also greatly improved. The original H.264 interpolation filter uses a 6-tap filter in the half of the brightness part. Quarter uses a bilinear interpolation filter, while the HEVC interpolation filter uses an 8-tap filter for half of the luminance part, a 7-tap filter for quarter, and an interpolation filter The operation time of HEVC is twice as long as that of H.264, and HEVC uses a fixed interpolation filter designed according to the signal processing theory, assuming that the video signal is an ideal low-pass signal. However, the video signal is not necessarily low-pass and not fixed. In recent years, deep learning-based methods have been widely used, and have achieved significant results in image and video processing, and this paper used support vector machine first divides the training data into four subgroups, and trains the model for four subgroups respectively. According to the message theory, the more side messages can effectively help the training, this paper uses SVM Features Mask and residuals as our side messages, and form our two-input model and three-input model. For the three-input model, BDBR can be reduced by 1.33% compared with HEVC, and the encoding time is saved by 13.39%, but the encoding can be saved The time is because Liu[18] support vector machine coding unit fast decision algorithm, in fact, the convolutional neural network increases the coding time, so we proposed an idea to use the simplest bilinear interpolation filter, and Finally, we use the proposed architecture to make an improvement. In terms of the three-pass architecture, BDBR can be reduced by about 1.09%, and encoding time can be saved by 20.64%.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	221	View/Open

社群 sharing

Loading...