Multimedia Systems Computer Networks and Communications
Reciprocal multimodal interaction and global semantic calibration for visible-depth-thermal salient object detection
Abstract
Visible-Depth-Thermal (VDT) salient object detection (SOD) combines complementary appearance, geometric, and thermal cues to locate salient objects reliably in complex environments, especially when individual modalities are degraded. However, existing VDT SOD methods remain vulnerable to heterogeneous cross-modal noise, while their top-down decoding processes often lack effective constraints for preserving global semantic consistency. To address these issues, we propose a Reciprocal Complementary and Global-Prior-Guided Network (RCGPNet), which improves detection reliability under heterogeneous modality degradation through closed-loop complementary interaction, adaptive multi-scale aggregation, and global semantic calibration. Specifically, a looped interaction mechanism establishes bidirectional information exchange among the three modalities, suppressing modality-specific interference and reducing noise accumulation caused by unilateral guidance. A region-aware pyramid strategy then evaluates contextual responses from different receptive fields and adaptively filters background interference. During decoding, decoupled cross-modal global priors calibrate spatial and channel representations across multiple levels, thereby enhancing object localization, structural consistency, and boundary precision. Extensive experiments on the VDT-2048 benchmark show that RCGPNet achieves the best performance on seven of nine evaluation metrics and remains competitive on the other two. Further evaluations on visible, depth, and thermal challenge subsets demonstrate its robustness against different modality-specific degradation conditions. These findings show that reciprocal noise suppression, adaptive context selection, and globally constrained decoding provide reusable methodological insights for reliable multi-modal visual computing.