| 摘要: | 在現今科技與人們生活密不可分的時代，高解析度的影像已經成為人們的日常需求。為了因應高解析度的影像，高效率視訊編碼能夠比上一代的視訊壓縮標準高出了兩倍的壓縮率，這是因為HEVC在影像壓縮技術中使用編碼單元、預測單元、轉換單元以及量化等方式，而在這影像壓縮過程中，為了降低傳輸資訊，使用量化參數導致影像的失真。所以本論文使用卷積神經網路的方式來對於失真影像進行補償，並且引入機器學習中的支持向量機，透過支持向量機來將卷積神經網路的訓練資料集進行分類，而在此提出兩種不同分類方式的主題，一個是利用支持向量機模型來分類，另一個則是使用支持向量機中rhoe的特徵來分類，將訓練資料集分成絕對簡單、相對簡單、相對複雜以及絕對複雜的訓練資料集，而這些特性集中的訓練資料集，在分別使用卷積神經網路去訓練以及優化影像；此外也將支持向量機應用於HEVC編碼端來進行編碼單元快速決策，以節省編碼時間，在畫面內預測中的實驗結果顯示，主題一對於影像品質平均提升0.254 (dB)左右的BDPSNR，並且節省14%左右的編碼壓縮時間，而主題二對於影像品質則是平均提升0.253 (dB)左右的BDPSNR，並且節省15%左右的編碼壓縮時間。除此之外也提出將支持向量機中的特徵復用於卷積神經網路方式，透過將支持向量機中變異數、平均值以及低頻交流值作成SVM Features Mask並引入到網路模型中，使模型預測更加精準，在HEVC畫面內預測中的實驗結果顯示，對於影像品質平均提升0.272 (dB)左右的BDPSNR。;In today′s era where technology is inseparable from people′s lives, high-resolution images have become people′s daily needs. In order to cope with high-resolution images, High-efficiency video coding can achieve a compression rate that is two times higher than the previous generation video compression standards. This is because HEVC uses coding units, prediction units, conversion units, and quantization in image compression technology. In this image compression process, in order to reduce the transmission information, the use of quantization parameters leads to distortion of the image. Therefore, this paper uses the convolutional neural network to compensate for the distorted image, and introduces support vector machines in machine learning. Through the support vector machine to classify the training data set of the convolutional neural network, it is proposed here. Two different classification themes, one is to use the support vector machine model to classify, the other is to use the characteristics of the support vector machine rho to classify, the training data set is divided into absolutely simple, relatively simple, relatively complex and absolutely complex The training data set, and the training data set in these feature concentration, respectively, use convolutional neural networks to train and optimize the image; in addition, the support vector machine is also applied to the HEVC encoding side to quickly make coding unit decisions to save coding time, The experimental results in the intra prediction show that Theme 1 improves the image quality by an average of BDPSNR of about 0.254 (dB) and saves about 14% of the encoding compression time, while Theme 2 improves the image quality by an average of BDPSNR of about 0.253(dB), and Save about 15% of encoding compression time. In addition, it is also proposed to reuse the features in the support vector machine for the convolutional neural network. By making the variance, average and low-frequency AC value of the support vector machine into the SVM Features Mask and introducing it into the network model, the model prediction is more accurate. Experimental results in HEVC intra prediction show that the image quality is improved by an average of BDPSNR of about 0.272 (dB). |
