Visible and Low-Resolution Infrared Images Fusion Detection Under Low-Visibility Conditions

Single-modality object detection often suffers from significant accuracy degradation under low visibility, while multi-modality fusion using high-resolution infrared cameras is expensive. To address these issues, we construct a pedestrian and vehicle dataset comprising paired visible and low-resolution infrared images under various low-visibility scenarios. Utilizing this dataset, we propose an efficient dual-stream detection network named Filtering Difference Fusion Network (FDF-Net). FDF-Net integrates a Median Filtering Module to suppress noise and artifacts caused by infrared image upsampling. A Difference Enhancement Module is designed to facilitate cross-modal feature interaction and enhancement at multiple scales. Finally, an Adaptive Fusion Module utilizes spatial and channel attention mechanisms to dynamically weight and fuse the multimodal features. Experimental results demonstrate that FDF-Net achieves an mAP50 of 84.8% and an mAP50:95 of 55.3%, outperforming the baseline by 4.4% and 3.2%, respectively. The proposed method significantly surpasses mainstream single-modality and fusion algorithms, providing a robust and cost-effective solution for visible and infrared image fusion detection.