LB-CIM: Low-Bandwidth Cross-Modal Interaction for Multimodal Visible-SAR Object Detection

Visible and synthetic aperture radar (SAR) sensors provide heterogeneous yet complementary observations for remote-sensing object detection. However, effective multimodal sensing fusion remains challenging because SAR features may contain locally unstable responses, while dense cross-modal interaction can introduce redundant information exchange and unnecessary computation. This paper presents LB-CIM, a low-bandwidth cross-modal interaction module for paired VIS–SAR object detection. LB-CIM adopts a three-stage design. First, multi-scale consistency denoising (MCD) estimates local SAR instability from cross-scale discrepancies and suppresses inconsistent responses before fusion. Second, differential low-bandwidth cross-attention (DLBCA) performs bidirectional interaction using queries at the unreduced interaction-grid resolution and spatially reduced key–value representations, while differential values and difference-aware gating emphasize non-redundant cross-modal information. Third, consensus-complementary enhancement (CCE) separates shared and complementary cues and adaptively reorganizes them to form the fused representation. LB-CIM is inserted at multiple feature levels of a dual-branch detector. Experiments on OGSOD-1.0 and SpaceNet6-OTD-Fog demonstrate consistent improvements over single-modal and multimodal baselines. On OGSOD-1.0, the proposed model achieves 95.1% mAP50 and 56.8% mAP50:95, improving the YOLO11-RGBT baseline by 4.8 and 4.7 percentage points, respectively. On SpaceNet6-OTD-Fog, it reaches 70.6% mAP50 and 38.5% mAP50:95, corresponding to gains of 2.7 and 2.0 percentage points. Ablation experiments confirm the complementary contributions of MCD, DLBCA, and CCE, while evaluations on YOLOv5n, YOLOv8n, and YOLO11n-based detectors show that the gains are maintained across different YOLO architectures. The complete model contains 4.6 M parameters and requires 10.4 GFLOPs, providing a favorable balance between detection accuracy and computational complexity. Qualitative comparisons further show fewer missed detections and false positives under diverse target densities and scene structures. These results demonstrate that LB-CIM provides an effective and computationally efficient solution for multimodal sensing representation learning and VIS–SAR information fusion in remote-sensing vision tasks.