Semi-supervised salient object detection via boundary-aware fusion and dynamic pseudo label optimization

Abstract

Semi-supervised learning has shown great potential in salient object detection (SOD) by reducing the reliance on pixel-level annotations. However, existing methods still face challenges in pseudo label generation and feature fusion. Feature fusion in the decoding stage often lacks explicit boundary modeling, leading to blurred and discontinuous predictions around object edges. Fixed-threshold selection strategies also struggle to adapt to variations in saliency map quality during training. To address these issues, we propose a semi-supervised SOD method combining boundary-aware fusion with dynamic-threshold pseudo label optimization. The boundary-aware fusion framework integrates three modules: Diverse Feature Aggregation, Explicit Boundary Guidance, and Guided Fusion Refinement, which collaboratively fuse semantic features, detailed features, and edge priors. Meanwhile, a dynamic thresholding mechanism adaptively adjusts pixel-level pseudo label selection, enhancing the use of unlabeled data. Experiments on multiple benchmark SOD datasets demonstrate that our method improves boundary sharpness and overall detection accuracy while maintaining training stability, outperforming existing semi-supervised SOD methods.