DAN-YOLO-L: a lightweight YOLO approach for efficient infrared small object detection from UAV perspectives

Small object detection in drone aerial photography faces challenges such as small scales, inconspicuous features, and complex backgrounds. To address these issues, this paper proposes an improved detection model based on the YOLO framework. First, a multi-channel feature extraction module-Deconvolutional Network combined with the C3 module (deconv-c3k2)-is designed to enhance feature extraction and multi-scale representation capabilities. Second, an enhanced Auxiliary Head detection module is introduced to improve feature interaction and collaboration across different levels. Concurrently, the NWD-Inner-CIoU loss function is adopted to mitigate the impact of IoU on small target localization offset, thereby boosting detection accuracy. To meet real-time embedded deployment requirements, an L1 pruning strategy is employed to reduce the model parameter size. Experimental results demonstrate that the proposed method significantly outperforms baseline models on the HIT-UAV dataset, achieving 81.9% on mAP@0.5 and 51.4% on mAP@0.5:0.95, respectively. Inference speed increases by 8.32%, while parameter count decreases by 20.8%. Computational load was reduced by 0.36 GFLOPs, and model size shrinks to 79.2% of the original. Furthermore, experiments on the IRay Infrared Dataset validate the method's generalization capability. Overall, the proposed approach demonstrates distinct advantages in detection accuracy, real-time performance, and lightweight design, while exhibiting stability and practical value.