產生VOP之穩定演算法的研究

Abstract

自動產生視訊物件平面的一個困難點在於並非整個物件所在的區域都會持續的移動，因此未移動的區域無法與背景區隔，也造成了從各視框找出的物件可能有相當的差異。常見的解決方法為以先前記錄的分割結果做為現在視框分割的參考。過去的方法包括以像素為單位記錄移動狀態或以區域為單位做記錄再做區域配對。在這篇論文中，我們將比較過去方法的優缺點，並深入探討實作的細節。我們融合過去方法的優點，提出一個有效率且能產生較穩定的物件之記錄方法，並減小物件的部份區域因沒有移動而被劃分為背景的錯誤。最後，我們提出兩種濾波器使得產生的物件有平滑且自然的輪廓。實驗結果顯示我們提出的方法確實能產生較佳的視訊物件。

A difficulty in automatic video object plane (VOP) generation is that not all regions in an object move constantly, making those non-moving parts hard to be distinguished from the background. The generated VOP thus may vary significantly from different frames. A common solution is referring to previous segmentation results in the current segmentation. Existing algorithms include recording the moving state with the granularity of pixels, or of regions as a whole followed by region matching. In this thesis, we compare the advantages and disadvantages of these existing algorithms, as well as study their implementation details. We combine the advantages of these existing algorithms to propose an efficient and stable memory mechanism. We also propose a method to reduce errors due to non-moving regions in the object. Finally, we propose two filters that generate smooth and natural object contour. The experimental results show our proposed methods can generate better video objects.