GPONet: A Two-Stream Gated Progressive Optimization Network for Salient Object Detection

The salient object detection task is to locate and detect salient regions in images, which is widely applied in various fields. In this paper, we propose a gated progressive optimization network (GPONet) for salient object detection. Firstly, to extract salient regions more accurately, we design a multi-level feature fusion module with a gating mechanism named gate fusion network (GFN). GFN focuses on the semantic information of high-level features as well as the detailed information of low-level features, enabling purposeful delivery of high-level features to low-level features. The gate fusion unit (GFU) is also able to maintain valid information and suppress redundant information in the fusion process. Secondly, while some existing methods have shown that the additional edge supervision can facilitate salient object detection, edge pixels are often much less common than non-edge pixels, leading to the challenge of class imbalance. To overcome this issue, we introduce detail labels that provide additional internal details as a supplementary supervisory signal. Combining these labels with proposed Detail Perception Loss enables our network to learn edge information of salient objects more effectively. To complement each other and guide information exchange between the two branches, we propose a cross guide module (CGM) to control the information flow transfer between them. Finally, we develop a simple and efficient attention fusion strategy to merge the prediction maps of the two branches to generate the final salient prediction map. Extensive experimental results validate that our method reaches optimal or comparable performance on several mainstream datasets. The code of GPONet is available from https://github.com/antonie-z/GPONet