| 摘要: | 近幾年來Covid-19病毒噬虐全球，已造成5億多人口染疫，更造成600多萬人死亡，且人數還在攀升中；台灣面臨這次疫情也造成好幾百萬人確診需要居家隔離，為了要降低染疫風險，學校改成線上教學，大部分的公司避免群聚也改成線上開會，受限於使用者所觀看的畫面解析度、系統處理能力、儲存容量及網路頻寬，傳送高畫質或超高畫質影像壓縮技術就變得非常重要，如何在有限的資源下避免造成畫面延遲或失真且同時可以將聲音完整傳達到對方是目前很熱門的研究議題。
新一代的HEVC(High Efficiency Video Coding)壓縮標準比起上一代H.264在相同畫面品質雖然可以達到兩倍的壓縮效率，但是編碼運算非常複雜且造成影像不可逆的失真，利用解碼端的後處理將失真補償回來是其中一種方法。本篇論文接續後處理的研究，利用像素與像素之間空間域的相關性提取影像特徵，包括高斯遮罩及殘差演算法，並參考消息理論中引入多個側面消息結構。在本篇論文中，我們提出以VDSR(Very Deep Super Resolution)及EDSR(Enhanced Deep Super-Solution)兩種CNN(Convolutional Neural Networks)模型為主架構，探討在不同的CNN架構下畫面間預測對於編碼後處理的效能評估，以不同演算法特徵進行CNN模型訓練，影片測試結果以VDSR為主架構的CNN，雙輸入及三輸入影像品質都有獲得明顯的改善，而以EDSR主架構的CNN測試結果，SVM分群的結果也明顯優於前者，BDBR減少0.381%~0.701% ，BD-PSNR提升0.037~0.171db。特別注意的是兩種架構在碼率伯仲之間時，其BD-PSNR卻可以提升近一倍。
;The covid-19 virus has spread all over the world in recent years. It has infected 500 million people and killed more than 6 million people, and the number is still rising; Taiwan is facing the epidemic, which has also caused tens of thousands of people to be diagnosed and need to be isolated at home. In order to reduce the risk of infection, schools have changed to online teaching; most of all the companies have also changed to online meetings to avoid gatherings. Limited by the screen resolution, system processing capability, storage capacity, and network bandwidth viewed by the user, it becomes a very important issue in transmitting high-quality or ultra-high-quality image compression technology. Reduce latency or distortion of the image and complete transmission of the sound to the other party at the same time is a very hot research topic.
The new generation of HEVC (High Efficiency Video Coding) compression standard can achieve 50% the compression efficiency compared to the previous generation H.264 in the same image quality. But the encoding operations are very complicated and cause irreversible image distortion. Distortion compensation is one of the methods. This paper continues the post-processing research, using the spatial domain correlation between pixels to extract image features, including Gaussian Mask and Residual algorithms, and enrolls the information theory in Introduces multiple side information structures. In this paper, we propose two CNN (Convolutional Neural Networks) models, VDSR (Very Deep Super-Resolution) and EDSR (Enhanced Deep Super-Solution) as the main architecture, to discuss the effect of Inter Prediction on encoder post-process under different CNN architectures. In the evaluation of processing efficiency, the CNN model is trained with different algorithm features. The image test results of the CNN with VDSR as the main structure have significantly improved the image quality of dual-input and triple-input, Forthmore, the test results of the CNN with the EDSR main structure, The results of SVM (Super Vector Machine) classification are also significantly better than the former, BDBR is reduced by about 0.381%~0.701%, and BDPSNR is improved by about 0.037~0.171db. It is particularly noteworthy that when the two architectures are in the same bitrate, EDSR’s BD-PSNR can be nearly doubled. |
