Weakly-supervised Salient Object Detection with Label Decoupling Siamese Network

Abstract

Weakly-supervised salient object detection (SOD) does not require a lot of manually annotated pixel-level labels. To further improve the detection accuracy of weakly-supervised SOD, we introduce a label decoupling siamese network (LDSN), which implements weakly-supervised SOD by learning from scribble labels. Scribble labels are novel weakly-supervised labels, which can reduce the cost of manual annotating and provide good interaction for users. In previous methods, the use of scribble labels mainly focuses on the salient object region. With the label decoupling siamese network, we make more adequately use of the scribble labels and the complementary relationship between salient object and background. We also propose a complementary self-supervised loss (CSS Loss) to realize a self-supervision. To improve the accuracy of the edge detection and interpretability of the network, we design a gradient priors weakly-supervised edge detection module (GWEM), which combines traditional image processing and deep learning methods. We compare the proposed network with 6 state-of-the-art SOD methods on 5 widely used datasets. The experiment results show that the performance of LDSN is superior to existing weakly-supervised SOD methods, and is compared with several pixel-level label supervised methods.