Cos-UMamba: Optimizing salient object detection with cosine scanning and bias-corrected feature fusion in optical remote sensing images

Salient object detection in optical remote sensing images (ORSI-SOD) is a critical task with wide-ranging applications, including environmental monitoring, urban planning, and disaster management. However, the effective fusion of local and global features remains a fundamental challenge in this field. While existing methods attempt to achieve feature complementarity through architectural innovations, the quadratic complexity of Transformers hinders their scalability, and traditional Mamba architectures suffer from static scanning limitations and lack dynamic adaptability. Moreover, representational bias in heterogeneous feature fusion is frequently overlooked, reducing the reliability of detection outcomes. To address these challenges, we propose Cos-UMamba, a novel hybrid framework that integrates bias correction mechanisms with a dynamic omni-directional cosine scanning strategy. This approach enables global long-range modeling of complex topological structures while effectively mitigating feature fusion bias through a K-nearest neighbor (KNN)-based graph construction. By eliminating interference from non-salient regions, the proposed model significantly enhances feature representation. Extensive evaluations conducted on standard ORSI-SOD datasets, including ORSSD, EORSSD, and ORSI-4199, demonstrate the superior performance of Cos-UMamba across multiple metrics such as mean absolute error (MAE) and F-measure. These results validate its capability to advance the accuracy and robustness of salient object detection in diverse remote sensing scenarios, offering a robust tool for tackling real-world challenges in the field. The source code and dataset will be available onhttps://github.com/darkseid-arch/Cos-UMamba.