DC 欄位 |
值 |
語言 |
DC.contributor | 通訊工程學系 | zh_TW |
DC.creator | 范聖敏 | zh_TW |
DC.creator | Sheng-Min Fan | en_US |
dc.date.accessioned | 2020-1-17T07:39:07Z | |
dc.date.available | 2020-1-17T07:39:07Z | |
dc.date.issued | 2020 | |
dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=106523021 | |
dc.contributor.department | 通訊工程學系 | zh_TW |
DC.description | 國立中央大學 | zh_TW |
DC.description | National Central University | en_US |
dc.description.abstract | 視訊編碼壓縮標準為高效率視訊編碼(High Efficiency Video Coding, HEVC)此壓縮編碼比起H.264擁有更高的編碼壓縮效率,其應用範圍也可以達到4K、8K的影像。HEVC在切割影像技術上使用四分樹(Quad-Tree)的編碼方式,將編碼樹單元(Coding Tree Unit, CTU)藉由壓縮編碼運算,CTU包含著四種不同的深度,深度最淺的影像失真較少;反之深度最深的影像失真較多,而這四種深度組成比例,也隨著量化參數(Quantization Parameter, QP)的不同有影響。而我們要利用卷積神經網路(Convolutional Neural Network, CNN)的方式來優化其影像品質,採用殘差網路的訓練方式,將訓練到的影像殘差值再補上失真影像,來達到強化影像的效果。而就如同上述所提及的量化參數影響CTU深度分布,其也會間接影像到神經網路對於影像品質優化的改善程度,於是我們利用支持向量機應用在CTU的快速演算法,作為我們CTU的分類器,將CTU分成簡單紋理與複雜紋理,再分別使用卷積神經網路優化其影像性能。我們使用卷機神經網路優化使用與未使用支持向量機作為CTU的分類器的兩種例子,發現有加入了支持向量機分類器對於卷積神經網路的訓練有四大優點,一是可以降低卷積神經網路的訓練資料量;二是可以多降低1%左右的BDBR;三是可以節省16%左右的編碼壓縮時間;四是對於特殊影像也有非常良好的效果。 | zh_TW |
dc.description.abstract | The video coding compression standard is High Efficiency Video Coding (HEVC). This compression coding has higher coding compression efficiency than H.264, and its application range can also reach 4K and 8K images. HEVC uses the Quad-Tree coding method in technology of cutting image. The coding tree unit (CTU) is compressed and calculated. The CTU contains four different depths and the lowest depth image distortion. Less; on the other hand, the image with the highest depth has more distortion, and the proportions of these four depths also have an effect with the quantization parameter (QP). We use the Convolutional Neural Network (CNN) method to optimize image quality, and use the training method of the residual network to add the distorted image to the residual value of trained image to achieve enhancement. Just as the above-mentioned quantization parameters affect the CTU depth distribution, it will also have different improvement in image quality optimization by the neural network. Therefore, we use support vector machines (SVM) to apply the fast algorithm of CTU as our CTU. The classifier classifies the CTU into simple textures and complex textures, and then uses a convolutional neural network to optimize its image performance. We use a convolutional neural network to optimize two examples of classifiers with and without support vector machines as CTUs. We find that adding a support vector machine classifier has four major advantages for the training of convolutional neural networks. First, it can reduce the amount of training data for convolutional neural networks; Second, it can reduce BDBR by about 1%; Third, it can save about 16% of encoding compression time; Fourth, it also has very good effects on special images. | en_US |
DC.subject | HEVC | zh_TW |
DC.subject | 改善編碼性能 | zh_TW |
DC.subject | 支持向量機 | zh_TW |
DC.subject | 卷積神經網路 | zh_TW |
DC.subject | 分散式編碼 | zh_TW |
DC.subject | HEVC | en_US |
DC.subject | Improved Coding Performance | en_US |
DC.subject | SVM | en_US |
DC.subject | CNN | en_US |
DC.subject | Distributed Coding | en_US |
DC.title | 一種應用於HEVC解碼端之深度學習架構的研究 | zh_TW |
dc.language.iso | zh-TW | zh-TW |
DC.title | Study of A Deep Learning Architecture For HEVC Decoder | en_US |
DC.type | 博碩士論文 | zh_TW |
DC.type | thesis | en_US |
DC.publisher | National Central University | en_US |