- 學位論文
透過近似最鄰近搜索達成視訊超解析度之演算法與實作
Fast Video Super Resolution via Approximate Nearest Neighbor Search : Algorithm and Implementation
摘要
圖片超解析度技術最近幾年吸引較多研究者的目光，但事實上影片超解析度在實際應用上比圖片超解析度技術重要許多。在這篇論文中，我們提出一個基於外部資料庫的快速影片超解析度技術，並使用趨近最鄰近搜索法來達成快速搜尋，並且利用影片在時域上較一致的特性來加速花在找最靠近的群來放大低解析度的小圖成為高解析度的小圖。 我們更進一步推展所提出的這個基於外部資料庫的影片超解析度演算法到GPU上，並且發現兩個問題，包括了搜尋的準確度問題，和一個區塊中的執行緒(thread)彼此間不一致的問題，這會導致GPU的運算效率降低。此外，我們也提出了一個解決方案能夠解決這兩個問題，我們命名它做球環(spherical ring)，它同時提高了執行緒的一致性並且增加了搜尋的準確度。 我們所提出的球環趨近最鄰近搜索法較我們使用過，並且是現在最好的趨近最鄰近搜索法RIANN快2~3倍。我們所提出的最後版本的基於外部資料庫的影片超解析度演算法藉由球環趨近搜索法達到了放大四倍到480P的解析度在速度平均30fps。
並列摘要
Image super resolution has gained more attention than video super resolution, but video super resolution is more important than image super resolution in practical. In this thesis, We propose an external based fast video super resolution technique via approximate nearest neighbor search, which exploits the characteristic of temporal coherency in video contents to speed up the time spent on searching for nearest cluster to super resolve low resolution patches to high resolution patches. We further extend the proposed external based video super resolution approach to GPU, and discover two problems from the proposed method mentioned above, including the searching accuracy and the problem of thread inconsistency in a thread block, which lowers the computing efficiency of powerful GPU. In addition, we propose a solution, named spherical ring, to both of the two problems, which rises the nearest neighbor search accuracy and lower the thread inconsistency simultaneously. The proposed spherical ring approximate nearest neighbor search is 2 to 3 times faster than the approximate nearest neighbor search algorithm, RIANN[4],which is used by the original proposed method. The final version of proposed external based video super resolution via spherical ring approximate search achieves 30 fps on average to 4 upsample frame to 480P resolution.