DNR-YOLO: an infrared small object detection algorithm for UAV-based imagery

Abstract To address the issues of feature attenuation, blurred boundaries, and background interference in the detection of small targets in infrared images captured by unmanned aerial vehicles (UAVs), a lightweight enhanced feature fusion detection algorithm, DNR-YOLO, based on the YOLOv11 architecture, is proposed. Firstly, a Depthwise Separable Enhanced Attention Module (DSEAM) is designed, which integrates channel attention and spatial texture enhancement mechanisms, and introduces a temperature regulation strategy and depthwise separable convolution to strengthen the contribution of shallow features to the representation of deep targets, effectively alleviating the problem of feature attenuation for small targets. Secondly, a GMM-NWD composite loss function is constructed, which models the positional uncertainty of target boxes based on Gaussian mixture models and combines the normalized Wasserstein distance to optimize boundary localization, thereby improving the detection accuracy of small targets in complex backgrounds. Finally, the improved S_RepViTBlock module is adopted, which expands the global receptive field through dilated convolution and integrates the DropBlock regularization strategy to enhance the discrimination ability for targets with blurred boundaries. Experimental results show that on the HIT-UAV enhanced dataset for UAVs, the DNR-YOLO algorithm achieves mAP50 and mAP50-95 of 90.09% and 65.3% respectively, while maintaining a real-time detection speed of 204 Frames Per Second (FPS), representing a 4.08% and 5.2% improvement over the baseline model YOLOv11n, and reducing the computational complexity measured in Giga Floating-Point Operations (GFLOPs) by 9.2%. The improved algorithm provides effective technical support for the detection of infrared small targets in practical engineering scenarios such as complex battlefield reconnaissance and power line inspection.