Abstract
Abstract
In remote sensing, small target detection is essential for applications like military target localization and traffic monitoring. However, small targets frequently lack adequate physical features, such as texture and shape, due to their limited pixel coverage. This leads to significant information loss during network propagation and decreases detection accuracy. This paper proposes an improved YOLOv12 model by integrating and adapting existing techniques to better address these issues, achieving a balance between detection accuracy and computational efficiency. To maximize feature extraction and improve the representation of small targets, a pinwheel-shaped convolution (PConv) module is integrated into the shallow feature extraction layers. The fixed-weight loss function in the conventional YOLO framework is replaced with a scale-based dynamic loss function (SBDL), which improves training stability and detection accuracy for targets of different sizes. Furthermore, the RFCBAMConv module is incorporated into the C3k2 component to form an enhanced C3k2-R structure, improving multi-scale feature representation. Additionally, the original large-object detection head (20
$$\times $$
20) is replaced with a 160
$$\times $$
160 head to better capture small-scale targets. Extensive ablation tests on the VisDrone-DET2021 and UAVDT datasets demonstrate that the modified YOLOv12 outperforms the original model in mAP@0.5 by 10.4%. The proposed modifications, based on adaptations and combinations of existing methods, achieve 83.6% mAP@0.5 while maintaining fast inference speed, providing an effective solution for small object detection with fewer parameters and lower GFLOPs compared to both heavyweight and lightweight models, making the model well-suited for real-time remote sensing and UAV-based monitoring applications.