DMD-YOLO architecture from infrared and visible images for object detection

In complex environments such as varying lighting, shadows, night-time conditions, smog, and background clutter, target detection remains a challenging task. Although target detection based on the dual-modalities of visible and infrared light can extract comprehensive features, the accuracy of target detection has not yet reached a satisfactory level. To address these issues, this paper proposes the DMD-Yolo target detection network. This network is a dual-input network for visible and infrared light, extracting features from each modality separately. To simultaneously consider the fusion of dual-modal feature information in both spatial and channel dimensions, a feature fusion module named DM module is designed and added to the backbone network to improve detection accuracy. To further enhance the model's ability to extract global features and considering the lightweight nature of the model, a CG block-C3 module is designed to replace some of the existing C3 modules. Experimental results show that the proposed DMD-Yolo network, based on the Yolo v5s architecture, achieves an average precision of 0.862 on the public dataset M3FD. Compared to the original single-modal Yolo v5 network, the map0.5 has increased by 15.1%, and the detection accuracy is improved compared to several existing dual-modal target detection networks.