影像分割技術在影像處理以及電腦視覺中都是一個重要且具挑戰性的問題,它分為兩類:多物件分割以及前景分割。前景物件分割的目標是將一張圖片中的前景從背景當中分離出來,這種技術可以運用到物件偵測或其他電腦視覺的應用當中。近來,已經有不少關於前景分割的研究,但是,這些研究通常是屬於監督式的方法,亦即需要使用者的一些互動才能得出結果,使得便利性不盡理想,例如圖形切割(Graph-Cut),使用者必須選取一部分前景和背景作為前景以及背景的種子,進而將整張圖片轉成圖形(Graph),並利用最小切割(min-cut)原理將圖片中的前景切割出來。 不同於傳統圖形切割的方法,在這篇論文中我們提出了一個非監督式的影像前景分割方法,以改善圖形切割的便利性。此方法利用了邊緣偵測以及基於邊緣分割的一些步驟,來擷取圖形切割所需的資訊,最後利用賽局理論的圖形切割方法來將影像分割成前景和背景。根據我們的實驗結果,相較於原本的賽局理論圖形切割,我們所提出的方法不僅不需要使用者的介入,結果也相當不錯。
Image segmentation is an essential and challenging problem in computer vision and image processing. It categorized into two categories, multi-label segmentation and figure-ground segmentation. The goal of figure-ground segmentation is to separate the object from background. It can be used in object detection or many other applications. Recently, a lot of methods have been proposed for solving figure-ground segmentation problems. However, most of them are supervised approaches. In other words, the procedures of those methods need some interactions of users. It makes those methods unfavorable. For example, Graph-Cut needs user to select a part of foreground and background to be foreground seeds and background seeds. A graph and min-cut theory are used to separate the foreground from the image. Different from traditional Graph-Cut approaches, we proposed an unsupervised figure-ground approach. It uses an edge-based method to grab required information for Graph-Cut. Then, we use game-theoretical Graph-Cut to divide the image to foreground and background. According to our experiment results, our method does not need user interaction and performs well compared with the previous Graph-Cut method.