向量分群方法在影像分割的應用

Abstract

本論文研究把影像前景從背景分離的影像分割方法。在光譜編織裡影像分割藉由最佳化包含拉普拉斯矩陣的目標函數求得。然而最佳化目標函數前景比例並不總是完整的前景物件。要得到更好的分割前景物件，必須考慮最佳的前景物件與次佳數個前景物件。 組合數個影像前景來產生出完整的影像前景，可以用非監督分類方法。本論文從圖論理論觀點研究光譜編織裡面的拉普拉斯矩陣，然後我們利用網路分群方法裡的模組值進行分群，分類後的社群對應了分割的影像前景。最佳化模組值將會轉成最佳化一組向量分群，我們提出一種向量分群的演算法，先藉由欲分類向量的正負號訊息尋找起始向量群組，然後執行非監督方法向量分割。 從經驗上的研究，向量分群能改進影像分割在測試影像上的結果。向量分群不但能從影像分辨出前景與背景，於前景物件也形成較少的分割區塊。這個新方法將會增強分割出前景和背景互相交織的影像。

This study investigates the segmentation of an image foreground from the background image. In the approach of image spectral matting, the segmentation of an image can be obtained by optimizing an objective function which contains matting Laplacian. However the optimized alpha matte of objective function is not always the entire foreground object. To obtain the better segmentation result of foreground object, the optimal alpha matte and the sub-optimal alpha mattes are all considered at the same time. The technique of unsupervised clustering can be applied to combine several foreground components into a complete foreground object. In this study, we investigate the matting Laplacian from the perspective of graph theory. Then we use the community detection method which is called network modularity to perform clustering. This detected community corresponds to the foreground component. Optimizing the modularity will turn out to be the vector partition problem. We propose an algorithm which finds the initial groups by the sign information of vectors to perform vector partition for unsupervised clustering Through empirical studies, the results of vector partition can improve the segmentation of test images. It can not only distinguish the foreground from the background, but also form less component regions of the foreground. This new approach will enhance the segmentation of the foreground object that is matted with background image components.