Lightweight Dual Correlation Variance Compensation for Salient Object Detection in Optical Remote Sensing Images

Salient Object Detection in Optical Remote Sensing Images (ORSI-SOD) confronts three fundamental challenges: extreme multi-scale object heterogeneity where targets span multiple orders of magnitude within single scenes, subtle inter-class boundaries with minimal contrast between objects and backgrounds, and computational scalability demands for high-resolution imagery processing. Existing approaches suffer from a critical limitation in that they fail to exploit the discriminative information encoded in inter-layer correlation variance, which contains essential boundary and transition cues that are systematically lost during conventional feature fusion processes. This paper introduces MambaCVC, a lightweight dual correlation variance compensation network that explicitly models and compensates for correlation variance patterns to enhance feature discriminability while maintaining linear computational complexity. Our approach centers on two synergistic modules: the Global Correlation Variance Compensation Module (GC-VCM) that selectively amplifies semantically meaningful long-range dependencies through correlation variance modeling across distant features, and the Local Correlation Variance Compensation Module (LC-VCM) that preserves fine-grained boundary details via variance-driven local enhancement mechanisms. To achieve computational efficiency, we leverage the VMamba backbone with minimal architectural adaptations optimized for ORSI-SOD. Comprehensive experiments on three public ORSI-SOD datasets demonstrate that MambaCVC achieves state-of-the-art performance with remarkable efficiency, requiring only 6.25G FLOPs and 7.26M parameters.