计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (12): 366-372.DOI: 10.3778/j.issn.1002-8331.2304-0397

• 工程与应用 • 上一篇    

改进YOLOv5n的管道DR缺陷图像检测方法

时亚南,陈志远,刘兆英,陈迎春,张婷,范效礼,苗锐,叶伟   

  1. 1.新疆维吾尔自治区特种设备检验研究院,乌鲁木齐 830011
    2.新疆特种设备检测技术研究重点实验室,乌鲁木齐 830011
    3.北京工业大学 信息学部,北京 100124
    4.北京工业大学 城市建设学部,北京 100124
  • 出版日期:2024-06-15 发布日期:2024-06-14

Improved YOLOv5n Pipeline DR Defect Image Detection Method

SHI Yanan, CHEN Zhiyuan, LIU Zhaoying, CHEN Yingchun, ZHANG Ting, FAN Xiaoli, MIAO Rui, YE Wei   

  1. 1.Xinjiang Uygur Autonomous Region Inspection Institute of Special Equipment, Urumqi 830011, China
    2.Xinjiang Key Laboratory of Special Equipment Testing Technology, Urumqi 830011, China
    3.Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
    4.Faculty of Architecture, Civil and Transportation Engineering, Beijing University of Technology, Beijing 100124, China
  • Online:2024-06-15 Published:2024-06-14

摘要: 近年来,数字射线成像技术(digital radiography, DR)由于其独有优势已被广泛应用于工业管道无损检测。为提高管道DR缺陷图像检测精度,提出了一种改进的YOLOv5n管道DR缺陷图像检测方法。该方法有两点贡献,针对目标检测网络中分类和回归两个任务存在冲突的问题,设计了任务独立解耦检测头,通过分别为两类任务构建独立的特征图实现解耦。为了缓解解耦检测头模块带来的参数量增加问题,引入了轻量化的深度可分离卷积替代标准卷积,在保证精度的同时,减少模型参数量。实验结果表明,在管道缺陷数据集上,该方法的mAP@0.5比YOLOv5n提高0.9个百分点。与YOLOv4、Faster-RCNN和SSD等其他几种目标检测模型的对比实验表明,该方法在mAP@0.5、参数量和计算量上都达到最优,有效提高了管道DR缺陷图像检测的性能。

关键词: 缺陷图像检测, 目标检测, 解耦检测头, 轻量化模型

Abstract: In recent years, digital radiography (DR) technology has been widely used for industrial pipelines defect detection due to its unique advantages. To improve the accuracy of pipeline DR defect image detection, an improved YOLOv5n is proposed. This method has two contributions. Firstly, to address the issue of conflicts between classification and regression tasks in target detection networks, a task independent decoupling detection head is designed, and decoupling is achieved by constructing independent feature maps for each type of task. Secondly, in order to alleviate the increase of parameters caused by the decoupling detection head module, a lightweight deep separable convolution is introduced to replace the standard convolution, which can effectively reduce the number of parameters while ensuring accuracy. The experimental results indicate that on the pipeline defect dataset, the mAP@0.5 of the proposed method is 0.9 percentage points higher than YOLOv5n. Comparative experiments with other object detection models such as YOLOv4, Faster-RCNN, and SSD show that the proposed method achieves superior performance in mAP@0.5, parameter quantity and computational complexity.

Key words: defect image detection, target detection, decoupling detection head, lightweight model