基于TAS-YOLO的道路表面缺陷检测
DOI:
作者:
作者单位:

武汉科技大学机械自动化学院 武汉 430081

作者简介:

通讯作者:

中图分类号:

TP391.4;U418;TN207

基金项目:

国家自然科学基金项目(52372395,51778509)资助


Road surface defect detection based on TAS-YOLO
Author:
Affiliation:

School of Mechanical Automation, Wuhan University of Science and Technology,Wuhan 430081, China

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    针对路面小缺陷检测准确率低、漏检率和误检率高且均匀分布缺陷类型数据集难以采集问题,本文提出一种基于YOLOv5s的TAS-YOLO改进网络模型方法。首先,在预测结果阶段采用特定任务的上下文解耦头,通过分离分类和定位任务,增强定位检测框的精度;其次,通过FPN结构将5个尺度的特征图输入解耦头进行预测,增强小目标的多尺度特征信息;最后,使用滑动窗口损失函数优化YOLOv5,提高难分类样本的检测准确率,且模型收敛效果更好。实验结果表明,TAS-YOLO算法提升了各类缺陷的平均检测精度,mAP50值达到91.4%,FPS值达到126,较YOLOv7l、YOLOv8s、YOLOv9c-gelan和Efficientdet等主流检测算法提高了检测的精度和效率。

    Abstract:

    This paper proposes an improved TAS-YOLO network model method based on YOLOv5s to address the issues of low accuracy, high missed and false detection rates, and difficulty in collecting uniformly distributed defect types datasets for detecting small road surface defects. Firstly, in the prediction result stage, a context decoupling head for a specific task is used to enhance the accuracy of the localization detection box by separating classification and localization tasks; secondly, by using the FPN structure to input feature maps of 5 scales into the decoupling head for prediction, the multi-scale feature information of small targets is enhanced; finally, use the silde loss function to optimize YOLOv5 and improve the detection accuracy of difficult to classify samples. The experimental results showed that TAS-YOLO algorithm improved the average detection accuracy of various defects, with mAP50 reaching 91.4% and FPS reaching 126, which improved the detection accuracy and efficiency compared with mainstream detection algorithms such as YOLOv7l, YOLOv8s, YOLOv9c-gelan and Efficientdet.

    参考文献
    相似文献
    引证文献
引用本文

李金涛,周兴林,尹雨飞,敖思铭.基于TAS-YOLO的道路表面缺陷检测[J].电子测量技术,2024,47(13):148-156

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2024-11-07
  • 出版日期: