基於結構差異性的影像分割

Abstract

我們提出一個新方法來處理影像分割問題，利用背景的結構一致性來區分前景。主要概念來自背景相減法。使用者只需要標示目標物體大略的邊框位置，我們的系統即可藉由最大化預測背景和真實背景的一致性，來找出物體的輪廓線。我們結合影像修補與前景分割的技術，使之融合為更有效的影像分割法。另外，我們也建立了一個新的影像資料庫，影像中的背景多為有結構性的物體，且前景物體顏色和背景顏色相似，難以判斷前景物體的輪廓線。利用我們的方法可有效切割出難以分辨的前景物體。

We introduce a novel approach to deal with image segmentation which takes into account the consistent structure of the backgrounds. The concept is from back- ground subtraction. Our method only needs users to specify their target objects by a bounding box. The system then finds the object contour by maximizing the consensus between the predicted background and the original image. We combine principles from image completion and foreground extraction approaches into a powerful uni?ed engine. Besides, a new and harder dataset is introduced with images which have structural background objects.