B2Net: Body and Boundary Learning for Salient Object Detection in Remote Sensing Images

Salient object detection (SOD) in remote sensing images (RSIs) has gained popularity recently. However, existing methods do not fully utilize edge information of salient objects. To address these challenges, we propose a Body and Boundary Learning Network (B2Net) to solve the problem of inaccurate edges in saliency maps. Specifically, the B2Net incorporates a body and boundary decoupling module to decompose salient objects into high-frequency boundary details and low-frequency central body features, overcoming the problem of uneven edge pixel distribution. In decoding, we design a multi-scale feature filtering module and a feature cross-fusion module to preserve local features and calibrate global information, thereby enhancing feature representability. Finally, the model is trained using a weighted fusion loss function to achieve accurate detection results. Experimental results show that the proposed algorithm significantly outperforms existing methods in SOD.