| 摘要: | 在影像品質不斷進步之下，人們對影像資料的需求量大幅增加。為了因應高解析度的影像，高效率視訊編碼(High Efficiency Video Coding，HEVC)能夠比上一代的視訊標準高出兩倍的壓縮率，這是因為高效率視訊編碼在影像壓縮技術中使用到編碼單元、預測單元、轉換單元以及量化等方式，而進行影像壓縮過程中，為了降低傳輸資訊，編碼使用到量化參數造成影像的失真。因此本論文解碼端使用卷積神經網路的架構進行反量化失真影像補償，而在此提出兩種卷積神經網路對於失真影像補償的主題，一個是CNN卷積神經網路對於各種影像品質優化的改善程度，另一個則是CNN卷積神經網路模型量化參數縮減。經過多次的實驗之後，在不影響原始影像的條件之下，解碼端透過CNN卷積神經網路模型只需要8個量化參數QP22,QP27,QP32,QP37,QP40,QP45,QP48,QP50取代原來QP31~QP51等31個量化參數且達到一樣的影像補償效果。;With the continuous improvement of image quality, people′s demand for image data has increased significantly. In order to handle high-resolution images, High Efficiency Video Coding (HEVC) can be twice as compressed as the previous generation of video standards. Because of HEVC uses coding units、prediction units、transfor units and quantization in image compression technologies. In order to reduce transmission information, encoding uses quantization parameters(QP) to cause image distortion. Therefore,decoding uses the convolutional neural network(CNN) architecture to perform inverse quantization of distortion image compensation in the end of this paper, and here are two topics of CNN compensation for distorted images, one is CNN for various image quality of the optimization, another is the reduction of the quantization parameters number of the CNN model. After many experiments, without affecting the original image, the decoder only needs 8 quantization parameters QP22, QP27, QP32, QP37, QP40, QP45, QP48, and QP50 to cover the original CNN model 31 quantization parameters such as QP31~QP51, and achieve the same image compensation effect.
Keywords: HEVC, Intra prediction, CNN, Compensate for the distorted image, Reduction the number of CNN Models |
