WS-YOLO:航拍视角小目标检测方法  

WS⁃YOLO:Small object detection method for aerial photography viewpoint

作　　者：王峣 蒋行国 秦海洋 黎明 刁豪杰 WANG Yao;JIANG Xingguo;QIN Haiyang;LI Ming;DIAO Haojie(School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644000,China;Artificial Intelligence Key Laboratory of Sichuan Province,Yibin 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000 [2]人工智能四川省重点实验室,四川宜宾644000

出　　处：《现代电子技术》2025年第5期68-74,共7页Modern Electronics Technique

基　　金：四川轻化工大学人才引进项目(2019RC12)。

摘　　要：航拍小目标检测中,由于图像背景复杂、目标尺度小、空间尺度变化大,导致无人机航拍出现漏检、误检等一系列问题。为提高无人机航拍检测准确性、实时性以及实现轻量化,文中基于YOLOv8n提出一种改进的无人机航拍视角小目标检测方法 WS-YOLO。首先,重构YOLOv8n网络结构,并增加160×160特征图对应预测头,以提高模型对小尺度目标检测的准确性与鲁棒性;然后,主干网络卷积层后嵌入SPD,防止细粒度信息的丢失并学习不太有效的特征表示,以提高对低分辨率图像的识别能力;最后,将CIoU损失函数替换为WIoU(v3)损失函数,以减轻图片质量低下对检测过程的影响。最终实验结果表明,环境、参数相同情况下,在VisDrone2019数据集与AI-TOD数据集上与原算法相比,WS-YOLO精确度能分别提升8.9%与7.7%,其参数量降低且FPS在合理范围,由此验证了提出的WS-YOLO在无人机航拍视角小目标检测中能够提高有效性。In the aerial photography small object detection,due to the complexity of the image background,the object scale is small,while the spatial scale varies greatly,which results in a series of problems such as leakage and false detection in UAV aerial photography.For this reason,an improved small object detection method WS-YOLO for UAV aerial photography viewpoint is proposed based on YOLOv8n,so as to improve detection accuracy and real-time performance,and realize the lightweight of the UAV aerial photography.Firstly,the YOLOv8n network structure is reconstructed and 160×160 feature maps corresponding to the prediction head are added,so as to improve the accuracy and robustness of the model for small-scale object detection.Then,the SPD is embedded after the convolutional layer of the backbone network,so as to prevent the loss of fine-grained information and learn less effective feature representations,so that its ability to recognize low-resolution images is improved.Finally,the CIoU loss function is replaced with the WIoU(v3)loss function,so as to mitigate the effect of low image quality on the process of detection.The final experimental results show that in the same environment and with the same parameters,the accuracy of the proposed WS-YOLO is improved by 8.9%and 7.7%on the VisDrone2019 dataset and the AI-TOD dataset,respectively,in comparison with that of the original algorithm.In addition,its parameters are reduced and its FPS is within a reasonable range.It is verified that the proposed WS-YOLO can improve the effectiveness in small object detection for the UAV aerial photography viewpoint.

关 键 词：无人机航拍 轻量化 YOLOv8n SPD WIoU VisDrone2019 AI-TOD 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]