Uncertainty-guided Siamese Transformer Network for salient object detection

Salient object detection acts with pivotal effect upon the analytical aspects of remote sensing images. However, the complexity of scene interference, diverse object shapes, and blurred boundaries in remote sensing images make it challenging for existing methods to accurately locate salient objects, predict clear boundaries, enhance feature representation, and effectively suppress background noise. To alleviate such issues, an Uncertainty- guided Siamese Transformer Network (USTNet) is proposed. This network constructs a Spatial-Semantic Complementary Feature Representation Module (SSRM) to effectively capture the global contextual information and fine-grained feature representations in remote sensing images, thereby better integrating multi-level and multi-scale features in these images and enhancing the expressive power of salient object detection. Meanwhile, an Uncertainty-guided Iterative Refinement Module (UIRM) based on foreground-background uncertainty modeling is designed, which dynamically estimates and rectifies the ambiguity and uncertainty between foreground and background features, gradually refining the representation and boundary description of salient regions. Extensive experiments demonstrate in which the suggested approaches have gained properties beyond the cutting-edge methods across multiple benchmark remote sensing image datasets. It significantly improves the accuracy of detecting objects that stand out in complicated contextual sceneries, boundary retention capability, and overall efficiency, showcasing its superior performance and broad applicability in complex remote sensing image analysis.