Architecture Optimization of YOLOv12 and Implementation of NWD-Based Hybrid Loss Function for Small Object Detection

Object detection in drone imagery faces challenges due to information loss during down-sampling and the prevalence of tiny objects. To address these issues, this paper proposes an improved YOLOv12s model optimized for aerial environments through structural redesign and loss function optimization. First, we reorganized the detection architecture to P2-P3-P4 by adding a high-resolution P2 head and removing the P5 head, while enhancing both the backbone and neck networks to preserve the spatial information of small objects. Second, we introduced a hybrid loss function combining NWD(Normalized Wasserstein Distance) and CIoU(Complete Intersection over Union) to mitigate gradient instability and ensure precise localization. Experiments on the VisDrone2019 dataset demonstrate that the proposed structural improvements are essential for maximizing the potential of the NWD loss function. The proposed model improved the AP50 from 28.0% to 37.1% and the Recall from 30.4% to 37.4% over the baseline, confirming its effectiveness in resolving the missed detection problem in drone surveillance.