基于Wasserstein距离的航拍图像小目标检测模型  

Small Target Detection Model in Aerial Images Based on Wasserstein Distance Loss

作　　者：蔡泽宇 刘远兴 李文炽 吴湘宁 杨翼 胡远江 CAI Zeyu;LIU Yuanxing;LI Wenzhi;WU Xiangning;YANG Yi;HU Yuanjiang(School of Computer Science,China University of Geosciences(Wuhan),Wuhan 430078,China)

机构地区：[1]中国地质大学(武汉)计算机学院,武汉430078

出　　处：《吉林大学学报(信息科学版)》2025年第1期65-76,共12页Journal of Jilin University(Information Science Edition)

基　　金：国家自然科学基金重点资助项目(U21A2013);智能地学信息处理湖北省重点实验室开放基金资助项目(KLIGIP-2018B14)。

摘　　要：针对无人机航拍具有多角度、大视角、大规模场景的特点,使无人机航拍图像存在小目标对象较多、背景复杂和特征提取困难的问题,提出了一种新的模型CA-NWD-YOLOv5(Coordinate Attention-Normalized Wasserstein Distance-You Only Look Once v5)。该模型以YOLOv5模型为基础,在头部网络添加了多尺度检测层以提取小目标特征,同时在骨干网络加入了CA注意力机制,避免模型忽略目标的位置信息。最后,使用归一化Wasserstein距离损失函数代替基于交并比的损失函数,加强了模型对微小目标的敏感程度。在VisDrone2019数据集上的实验表明,相比改进前的YOLOv5模型,CA-NWD-YOLOv5模型可有效提升无人机航拍图像中小目标的检测精度,改进后算法的mAP_0.5达到了50%,可以有效应用于航拍小目标的检测。UAV(Unmanned Aerial Vehicle) aerial photography,characterized by multi-angle,large field of view,and large-scale scenes,often results in images with numerous small objects,complex backgrounds,and difficult feature extraction.To address these issues,a new model,CA-NWD-YOLOV5(Coordinate Attention-Normalized Wasserstein Distance-You Only Look Once v5) is proposed.Based on the YOLOv5 model,a multi-scale detection layer is added to the head network to extract the features of small targets.It also incorporates a CA attention mechanism into the backbone network to prevent the model from overlooking target location information.Lastly,the normalized Wasserstein distance loss function replaces the loss function based on intersection ratio,enhancing the model's sensitivity to small targets.Experiments on the VisDrone2019 dataset demonstrate that,compared to the improved YOLOv5 model,the CA-NWD-YOLOv5 model can effectively enhance the detection accuracy of small and medium-sized targets in UAV aerial photography images.The mAP_0.5 of the improved algorithm reaches 50%,proving its effective application to the detection of small targets in aerial photography.

关 键 词：航拍图像 小目标检测 注意力机制 Wasserstein距离 

分 类 号：TP319.4[自动化与计算机技术—计算机软件与理论]