Co-saliency Detection via Weakly Supervised Learning

Abstract

A new weakly supervised learning (WSL) approach has been proposed in this paper for co-saliency detection. Most co-saliency detection algorithms in literature extract only a few parts of the salient object. Some algorithms require large training datasets. They generally exhibit poor performance in identifying salient objects having multiple colors. The proposed WSL approach is aimed at ameliorating these limitations. The WSL approach involves the careful refinement of the foreground and the background of a single image in order to train a support vector machine to classify the regions of the common salient object in a set of related images. The WSL method has been evaluated on imagepair and iCoseg, public domain benchmark datasets. The WSL method exhibits superior co-saliency detection performance to several state-of-the-art methods.