以邊緣偵測為基礎的高效率強健式視訊物件分割技術

Abstract

本論文提出一個新的視訊物件分割演算法。這個視訊物件分割演算法可區分為兩部分：初始背景的建立與物件的追蹤。在第一個部分，我們根據些許連續的影像建立出可信賴的初始背景，並且使用改善過的相連元件法(Modified Connected Component Method)將一張物件影像分割成許多相同灰階的區塊。然後，利用邊緣運算器找出物件的移動邊緣，再依照此資訊找出揭開背景(Uncovered Background)的區塊，最後更新初始背景。而在第二個部分，我們使用背景資訊和邊緣運算器追蹤新物件的邊緣，並透過改變偵測和背景預測的方法移除揭開背景的邊緣，進而抽取出完整的視訊物件。實驗證明利用背景資訊和邊緣資訊，我們可以有效地分割出精確的物件並且改善以往只用改變偵測(Change detection)作為視訊物件分割的缺點。

In this thesis, we propose a new video object segmentation algorithm. The video object segmentation algorithm consists of two major parts: initial background construction and object tracking. In the first part, we construct the reliable initial background in several consecutive frames and use modified connected component to partition an object image into many blobs with similar luminance. Then, we use edge operator to find the moving edge and use it to find the uncovered background blob. Finally, the initial background frame could be updated. In secondary part, we use background information and edge operator to find the moving edge of the object. Then, the uncovered background edge is removed by using the change detection method and background predictive method. Further, the perfect video object could be extracted. According to the experimental results, the proposed method combining background information and edge information can greatly improve the performance of precise object segmentation compared with the conventional change detection approaches.