Cross Domain Edge Detection Based Label Decoupling Salient Object Detection

Sida Zheng, Jichang Guo, Huihui Yue, Xiaowen Liu

Abstract: The edge of salient object plays an important role in salient object detection. In this paper, a cross domain edge detection based label decoupling salient object detection network (CDENet) is proposed to improve the accuracy of saliency detection and make more adequate use of the edge information in an image. CDENet decouples the original salient object detection labels, and is trained with original saliency maps and the decoupled labels. In order to make better use of edge information, a cross-domain edge detection module (CDED) is proposed, which fuses the features of RGB domain and HSV domain. In addition, we also design a cross self-supervised module (CSSM) for label decoupling network. CSSM uses the implicit association between body prediction and edge prediction to strengthen the information interaction between different branches of the neural network. Comprehensive experiments on 5 widely used data sets in salient object detection show that CDENet is superior to 6 state-of-the-art algorithms in several objective indicators. Experiment results show that the proposed approach is an effective saliency detection method. CDED and CSSM can accurately fuse features of different branches and improve the salient object detection accuracy.