Adversarial Learning Based Saliency Detection

Abstract

Image saliency detection has recently witnessed rapid progress due to deep convolutional neural networks. However, the typical binary cross entropy loss used in the networks by saliency detection is a pixel-wise loss, resulting in the independent prediction of the salient probability of each pixel. It raises the problem of spatial discontinuity of the predicted saliency maps. Many researchers try to solve this problem by using super-pixel segmentation, but it is complicated and time-consuming. In this paper, we propose an Adversarial Saliency Detection Network (ASDN) to enhance the spatial continuity of the saliency maps with two sub-networks which are saliency detection network and discriminator network, respectively. The aim of the discriminator is to distinguish the saliency maps predicted by the saliency detection network from the ground truth. In this way, the discriminator helps the saliency detection network to enhance long-range spatial continuity of the predicted saliency map. Our ASDN achieves the state-of-the-art performance on standard salient object detection benchmarks.