結合支持向量機與摺積神經網路以提升HEVC編碼效能之研究;A Combined Support Vector Machine and Convolutional Neural Network Architecture for HEVC

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/82853

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/82853

Title:	結合支持向量機與摺積神經網路以提升HEVC編碼效能之研究;A Combined Support Vector Machine and Convolutional Neural Network Architecture for HEVC
Authors:	張詠鈞;CHANG, YUNG-CHUN
Contributors:	通訊工程學系
Keywords:	HEVC;去區塊濾波器;支持向量機;畫面間預測;畫面內預測;移動向量;RDO;摺積神經網路;深度學習;HEVC;Deblocking filter;SVM;Inter Prediction;Intra Prediction;Motion Vector;NeighboringRDO;Convolutional Neural Network(CNN);Deep Learning
Date:	2020-01-17
Issue Date:	2020-06-05 17:24:38 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著科技的日新月異，人們對於高畫質的追求始終鍥而不捨，因此高解析度的顯示器及影像產品也就越來越多，而為了能夠有效壓縮高解析度中的龐大資料量，HEVC( High Efficiency Video Coding )使用許多方法來有效的降低位元率。然而為了更精進位元率以及畫質的表現，在畫面間預測上，我們應用支持向量機SVM ( Support vector machine )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、鄰近區塊深度資訊做為特徵(Feature)將一個CTU分類成只做深度0、深度0~1、深度0~2、深度0~3四種類別，以此略過特定深度的運算。預測單元以畫面間預測的移動向量值的資訊、Skip flag、鄰近區塊RDO資訊做為特徵(Feature)，判斷預測單元做完Inter2N×2N後是否需要提前中止，進而節省掉後續預測模式所需花費的運算時間，不僅如此，我們再結合近年來日益普及的摺積神經網路CNN ( Convolutional Neural Network ) 於HEVC中的環路濾波器( In-Loop filter )來提高畫質的表現。由於藉由SVM分類的圖片中有相似的性質來訓練神經網路，比起未分類深度可以達到更好的提升效果。最後結合兩種演算法與CNN來與HEVC進行比較。而在畫面內預測上，擷取原始畫面的資訊以及空間上的相關性做為特徵，把CTU分為只做深度0~2以及深度0~3兩種組別，依照輸入特徵給予SVM預測結果來判斷是否當前CU要提早略過或提早終止，並也在其HEVC中的環路濾波結合以分類結果所訓練出的CNN模型來提升影像品質。此研究不僅在HEVC畫面內預測上做改良，畫面間預測也有相當表現，各別依照不同SVM架構與匹配神經網路來達到提高影像的效果，於畫面內預測上我們能達到BD-PSNR (0.36 dB)、BD-BR (-6.2%)；畫面間預測上能達到BD-PSNR (0.25 dB)、BD-BR (-6.2%)甚至能減少6%的編碼時間。;With the rapid development of technology, People are always persistent in pursuing the high quality of video. Therefore, multimedia devices like monitors, players that have high resolution started rapidly increasing in numbers. In order to compress the significant increasing of data storage effectively, HEVC utilize multiple techniques to efficiently decrease bitrate. In inter-pridection, for the better effects, we proposed SVM-based fast inter CU ( Coding Units) depth decision algorithm and SVM-based fast inter PU mode decision algorithm to reduce the computational complexity. In SVM-based fast inter CU depth decision algorithm, we can skip certain depth by using SVM with features, including motion vector variance, CBF of merge mode, neighboring CU depth to classify a CTU into depth 0, depth 0~1, depth 0~2 and depth 0~3. In SVM-based fast inter PU mode decision algorithm, we use SVM with features, including motion vector variance, skip flag, the information of neighboring RDO to classify whether do early termination at 2N×2N. Besides it, we also combine CNN model with SVM in In-Loop filter of HEVC. CNN is a more and more popular technique wich can help us not only to recognize images or objects but enhance performance of portrait recently. So we can use the models to deal with the reconstruct images and thence enhance the quality of pictures. With the similar natures of blocks which SVM classes with, blocks in the same groups are trained together. Consequently, we get the models with different effects for distinct groups respectively and due to the relationship between the groups and the models, we can get the better performance than the results obtained by only using CNN without SVM. Finally, we combine two algorithms and CNN to compare with HEVC. Furthermore, in intra-prediction, by applying SVM with features consist of the CUs’ information and space relation, it can develop the criterion of early CU splitting and termination so that we can speed up intra-prediction by classifying a CTU into depth 0~2, depth 1~3. Again, we also use the classifications to train CNN model, and introduce it in deblocking filter on purpose to enhance the image performance. We improve effect on intra-prediction as well as inter-prediction, and both they can get eminent achievement. Our experiment results that the method surpasses mode (HM) with BD-PSNR (0.36 dB), BD-BR (-6.2%) on intra-prediction and BD-PSNR (0.25 dB), BD-BR (-6.2%) on inter-prediction which can even get 6% time saving compared to HM16.0.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	122	View/Open

社群 sharing

Loading...