Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (1): 110-121.DOI: 10.3778/j.issn.1002-8331.2304-0150

• Special Issue on YOLO Improvements and Applications • Previous Articles     Next Articles

Improved YOLOv5s UAV View Small Target Detection Algorithm

LIU Tao, GAO Yimeng, CHAI Rui, LI Zhengtong   

  1. 1.School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
    2.Department of Basic Teaching, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2024-01-01 Published:2024-01-01

改进YOLOv5s的无人机视角下小目标检测算法

刘涛,高一萌,柴蕊,李政通   

  1. 1.辽宁工程技术大学 软件学院,辽宁省 葫芦岛 125105
    2.辽宁工程技术大学 基础教学部,辽宁省 葫芦岛 125105

Abstract: The small target image from the UAV perspective has the characteristics of dense target distribution, unbalanced category and inconspicuous features, which leads to the problem of missed detection and false detection in the target detection task. To solve these problems, an improved YOLOv5s small target detection method is proposed to improve the accuracy and accuracy of target detection. First, it reclusters the anchor box to lock the detection area more accurately. Secondly, the backbone network structure is changed and convolution is added to the spatial pyramid pool layer to ensure that the detection target features are fully obtained. At the same time, the C3 module in the network structure is replaced with a lightweight SEC2f module that fuses the channel attention mechanism to improve the local feature acquisition ability of the network for small target detection. Finally, the features of the target area are extracted effectively by combining the decoupled detection head with the adaptive anchor frame calculation. Under the same parameters and environmental conditions, the detection accuracy on DOTA data set and VisDrone data set is improved by 6.1% and 5.2%, respectively, indicating the effectiveness of the improved method on small target detection tasks. The comparison experiment on voc2007+2012 public data set shows the universality of the improved algorithm.

Key words: YOLOv5s, clustering algorithm, SEC2f module, spatial pyramid pool, decoupling detection head

摘要: 无人机视角的小目标图像具有目标分布密集、类别不均衡以及特征不明显的特点,导致目标检测任务中出现漏检、误检的问题。针对这些问题,提出一种改进YOLOv5s小目标检测方法,以达到提高目标检测准确率与精确度的目的。重新聚类锚框,更精确地锁定检测区域。更改骨干网络结构,在空间金字塔池化层增加卷积,保证充分获取检测目标特征。同时,将网络结构中的C3模块替换成融合通道注意力机制的轻量级SEC2f模块,以提升网络对于小目标检测的局部特征捕获能力。融合解耦检测头,结合自适应锚框计算,有效提取目标区域的特征。在相同参数、相同环境条件下,在DOTA数据集上和VisDrone数据集上检测精度分别提升6.1%、5.2%,表明改进方法在小目标检测任务上的有效性;在公开数据集voc2007+2012上做通用性对比实验,结果表明改进算法具有通用性。

关键词: YOLOv5s, 聚类算法, SEC2f模块, 空间金字塔池化, 解耦检测头