ATS-YOLO：基于改进YOLO11的轻量化无人机小目标检测方法

ATS-YOLO: An Improved YOLO11-Based Lightweight Detector for Small Targets in UAV Imagery

摘要: 无人机目标检测面临小目标尺度小且密集、视角变化大与目标姿态多样，以及黑夜和强光等复杂光照环境带来的特征退化问题，同时受限于平台计算资源，使得检测精度、鲁棒性与实时性之间难以兼顾。为解决以上问题，本文提出了基于YOLO11的改进检测模型ATS-YOLO。首先，在主干网络中引入C2f-CA模块以增强特征表达能力，在检测头中引入TADDH结构以提升多尺度特征融合效果，并在训练阶段对边界框回归损失进行改进以强化对小目标的关注。通过上述设计，从特征提取、特征融合与损失约束三个层面实现了协同优化。实验结果表明，所提出组合ATS-YOLO在VisDrone2019数据集上相较于基线YOLO11模型在小目标检测性能上取得了明显提升，ATS-YOLO相比YOLO11n在mAP@0.5和mAP@0.5:0.95上分别取得了0.025、0.018的提升，在Params和Model size上分别减少了0.292 M和0.61 MB。此外，在对比实验上，改进模块的指标也显著优于其他算法。实验结果表明，ATS-YOLO在保持模型轻量化的同时实现了更优的检测效果，验证了该组合策略的有效性，为无人机场景下高效、鲁棒的目标检测提供了有效解决方案。

Abstract: UAV target detection is faced with the problems of small and dense small targets, large changes in perspective, diverse target poses, and feature degradation caused by complex illumination environments such as dark nights and strong light. At the same time, it is limited by platform computing resources, making it difficult to balance detection accuracy, robustness and real-time performance. In order to solve the above problems, this paper proposes an improved detection model ATS-YOLO based on YOLO11. Firstly, the C2f-CA module is introduced into the backbone network to enhance the feature expression ability, the TADDH structure is introduced into the detection head to improve the multi-scale feature fusion effect, and the bounding box regression loss is improved in the training stage to strengthen the focus on small targets. Through the above design, collaborative optimization is realized from three aspects: feature extraction, feature fusion and loss constraint. The experimental results show that the proposed combined ATS-YOLO has achieved a significant improvement in small target detection performance compared with the baseline YOLO11 model on the VisDrone2019 dataset. Compared with YOLO11n, ATS-YOLO has achieved 0.025 and 0.018 improvements on mAP@0.5 and mAP@0.5:0.95, respectively, and reduced 0.292 M and 0.61 MB on Params and Model size, respectively. In addition, in the comparative experiment, the index of the improved module is also significantly better than other algorithms. The experimental results show that ATS-YOLO achieves better detection results while maintaining the lightweight of the model, which verifies the effectiveness of the combined strategy and provides an effective solution for efficient and robust target detection in UAV scenarios.