- 學位論文
基於輕量權重方法之視頻影像超解析
A Study on Lightweight Based Method for Video Super Resolution
摘要
近年來，影像修復與增強任務在圖像處理界引起越來越多的關注，深度學習已廣泛應用於視頻超分辨率（VSR）。大多數 VSR 方法專注於實現更好的質量，深度神經網絡的設計也變得越來越複雜。而影像資料輸入數據比單張圖像資料大好幾倍，一般電視規範一秒的影片約有30個影格，如果神經網絡過於復雜，執行速度會更慢，消耗更高的內存負載，需要更高階的顯卡設備。 由於VSR的運算量相當繁重，大型實驗室用的顯示卡是相當高階的，為了讓低階電腦設備能夠做神經網路的訓練及推論，影像超解析的輕量化問題是值得被研究的。在本文中，我們提出了一種基於殘差密集塊的神經網絡稱為RDLNET。 在實驗中，與其他最先進的 VSR 方法相比，我們提出的輕量型VSR 方法減少了約 25% 的參數並保持幾乎相同的 SSIM。
並列摘要
In recent years, video restoration and enhancement tasks have attracted more and more attention in the image processing community, and deep learning has been widely used in video super-resolution (VSR). Most VSR methods focus on achieving better quality, and the design of deep neural networks has become increasingly complex. The input data of video data is several times larger than that of a single image data. Generally, there are about 30 frames per second of video in the TV standard. If the neural network is too complex, it will cause a higher memory load and require a higher-end graphics card device. Due to the heavy computing load of VSR, the graphics cards used in large laboratories are quite high-end. In order to enable low-end computer equipment to do neural network training and inference, the lightweight problem of video super-resolution is worth studying. In this paper, we propose a residual dense block based neural network called RDLNET. In experiments, our proposed lightweight VSR method reduces about 25% parameters and maintains almost the same SSIM compared to other state-of-the-art VSR methods.