一個快速動態影像預測區塊比對法

Abstract

在本論文中，我們提出了一個快速的動態影像預測區塊比對法。根據 一個區塊內圖像的灰階值差異，影像區塊可區分為平滑區塊與非平滑區塊 。對於平滑區塊，我們用少量的代表圖點來尋找最佳區塊的移動向量。對 於非平滑區塊，我們將一些不可能是最佳區塊的候選區塊予以踢除,再逐 一比對其它區塊,找出最佳的移動向量。這使得我們的方法能大量地降低 計算時間。我們使用三步搜尋法的搜尋策略，先以大範圍來搜尋最有可能 出現最佳區塊的區域，再逐步收斂至最佳區塊。實驗結果顯示，我們的方 法比著名的三步搜尋法快約一倍，且得到近似的影像品質。若與全域搜尋 法比較，我們的方法比它快約20倍，而影像品質並沒有很明顯的差異。 In this thesis, a fast block-matching algorithm(BMA) for motion estimation is proposed. According to the standard deviation of a block, the blocks areclassified into two classes －smooth blocks and non-smooth blocks. In a smoothblock, only a subset of pixels in a block is calculated to find the minimalmean absolute difference(MAD) blocks. In a non-smooth block, impossibleoptimal candidate blocks are rejected immediately and others are compared tofind the optimal moving vector . These classification and process lead toreduce the total computation time significantly. The algorithm followsThree-Step-Search (TSS) to search a most possible area widely first, then it steps to local minimal finally. Experimental results show that the performanceis improved not only with twice speed-up times than TSS approximately insimilar quality of image, but also more than 20 speed-up times than Full Search(FS)algorithm without significant difference in the quality of image.