Co-saliency Detection via Extremely Weakly Supervised Convolutional Neural Network

Abstract

A new computational intelligence approach called extremely weakly supervised learning (EWS) has been proposed for co-saliency detection in this paper. Most co-saliency detection algorithms in the literature are quite ineffective in highlighting the entire salient object. Further, they fail in providing high saliency values to some parts of the salient object. In addition, most learning-based algorithms require voluminous training datasets in order to enable the learning of the common object features. The EWS approach proposed in this paper is aimed at circumventing these limitations. The approach involves the training of a one-dimensional convolutional neural network with a careful refinement of foreground and background superpixels and the use of a single image in order to extract the common salient object from an image group. Experiments on Imagepair and iCoseg, public-domain co-saliency benchmark datasets, show that the EWS approach is quite effective in comparison with the state-of-the-art methods.