Automatic Foreground Initialization for Binary Image Segmentation

Foreground segmentation is a fundamental problem in computer vision. A popular approach for foreground extraction is through graph cuts in energy minimization framework. Most existing graph cuts based image segmentation algorithms rely on user’s initialization. In this work, we aim to find an automatic initialization for graph cuts. Unlike many previous methods, no additional training dataset is needed. Collecting a training set is not only expensive and time consuming, but it also may bias the algorithm to the particular data distribution of the collected dataset. We assume that the foreground differs significantly from the background in some unknown feature space and try to find the rectangle that is most different from the rest of the image by measuring histograms dissimilarity. We extract multiple features, design a ranking function to select good features, and compute histograms based on integral images. The standard graph cuts binary segmentation is applied, based on the color models learned from the initial rectangular segmentation. Then the steps of refining the color models and re-segmenting the image iterate in the grabcut manner, until convergence, which is guaranteed. The foreground detection algorithm performs well and the segmentation is further improved by graph cuts. We evaluate our method on three datasets with manually labelled foreground regions, and show that we reach the similar level of accuracy compared to previous work. Our approach, however, has an advantage over the previous work that we do not require a training dataset.