電機資訊學院: 資訊網路與多媒體研究所指導教授: 莊永裕潘冠綸Pan, Kuan-LunKuan-LunPan2017-03-062018-07-052017-03-062018-07-052015http://ntur.lib.ntu.edu.tw//handle/246246/275943從一連串觀測到的低解析度影像，合成一張高解析度影像的演算 法，稱之為多張影像超解析度。多張影像超解析度演算法的其中一個 關鍵在於，不同張低解析度影像之間，要找到精確的對應關係，使不 同張低解析度影像能相互對齊。 在論文中，首先運用光流法與單應性來對齊影像，它們性質並不相 同，並且分別在不同的狀況下，產生不理想的超解析度結果。我們延 伸單應性對齊，成為多重單應性的對齊方法。在原先單應性對齊方法 中，只會從兩張圖片找一個單應性矩陣。但是我們延伸成找多個單應 性矩陣，之後再對每個像素標記屬於哪個單應性矩陣以及進一步的修 正。 透過此方法，先前無法用單應性或光流法產生理想超解析度結果的 狀況能被克服，產生好的超解析度結果，並且時間與記憶體的使用上 都較光流法更有效率。除此之外，此方法也相當適用於，當大影像需 要做多次但每次只要做一小塊的超解析度的時候。Multi-frame image super resolution aims to produce one high resolution image from a set of observed low resolution images. One of key components to the success of super resolution algorithm is an accurate alignment between observed images. In this paper, we first show that two popular alignment methods—optical flow and homography alignment—have their own flaws and fail to produce good super resolution results in different cases. We propose a novel alignment method called multi-homography alignment for super resolution. This method is an extension of the homography alignment. In origin homography alignment, only a single homography is extracted from images. However, we extract multiple homographies from images in our alignment method and label each pixel to a proper homography to warp and refine the result. We show this method produces good super resolution results in previous cases that homography or optical flow cannot perform well and it is more efficient both in time and memory than optical flow. Finally, we point out that our alignment method is also very suitable for multiple times small region super resolution in large images.75636221 bytesapplication/pdf論文公開時間: 2016/8/17論文使用權限: 同意有償授權(權利金給回饋學校)超解析度對齊單應性光流法小區域super resolutionalignmenthomographyoptical flowsmall regionthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/275943/1/ntu-104-R02944017-1.pdf