HFDANet: A Hierarchical Feature and Dual-Branch Adaptive Refinement Network for Salient Object Detection in Optical Remote Sensing Images
Weijun Liu · Journal of Physics Conference Series · 2025
Abstract The complex background and scale variations of the same objects pose significant challenges, making them difficult to recognize in an ORSI image. To enhance feature representation capability and achieve refined object localization, this paper proposes an innovative network, HFDANet (Hierarchical Feature and Dual-Branch Adaptive Refinement Network), which effectively addresses these issues. The core of HFDANet is the DARE (Dual-Branch Adaptive Refinement Engine), complemented by the HFAIM (Hierarchical Feature Aggregation Integration Module). HFAIM integrates hierarchical feature fusion and aggregation to enhance semantic coherence and preserve fine-grained details of salient regions. The DARE module refines feature representations by effectively separating foreground from background and accurately localizing salient objects, resulting in superior detection performance even in challenging scenarios. Extensive experiments on the ORSSD, EORSSD, and ORSI-4199 datasets demonstrate that HFDANet significantly outperforms state-of-the-art methods in both robustness and detection accuracy for salient object detection tasks.