IVD-NET: An Adaptive Dual-Branch Network for Small Object Detection in Multimodal Remote Sensing Images

To address challenges such as drastic illumination changes, dense small objects, and frequent occlusions in remote sensing and UAV Images, this paper proposes an Infrared-Visible Dual-branch Detection Network (IVD-NET). Existing fusion methods often overlook cross-modal feature alignment and insufficient representation of small objects. To tackle these issues, IVD-NET employs a dual-stream architecture with three key innovations. First, an Adaptive Modality Fusion Module (AMM) comprising Cross-modal Correlation Attention (CCA) and Cross-modal Guided Pixel Attention (CGPA) is proposed to dynamically integrate complementary information from infrared and visible modalities, effectively handling lighting interference. Second, a Visual Parallel Aggregation (ViPA) module is designed to enhance feature representation for small objects through parallel multi-scale branches and frequency-domain attention. Finally, a Dual-Path Adaptive Detection Head integrating the Spatial Enhanced Attention Module (SEAM) is introduced to improve localization and classification accuracy for occluded targets. Extensive experiments on DroneVehicle and LLVIP datasets demonstrate that IVD-NET achieves state-of-the-art (SOTA) performance. Specifically, on the DroneVehicle dataset, it achieves an mAP50 of 83.08%, outperforming methods like YOLOv10 and MMDetection.