基于轻量化YOLO网络的实时X射线焊缝缺陷检测

Real-time X-ray Weld Defect Detection Based on Lightweight YOLO Network

X射线无损检测是进行焊缝焊接质量检查的重要方法,为了提高射线评片效率和准确率,本文提出了基于轻量化YOLO网络的实时X射线焊缝缺陷检测方法。首先,在分析焊缝射线检测图像特征的基础上,为提升焊缝图像缺陷检测精度与速度,设计了级联缺陷检测模型,先使用经过轻量化设计的焊缝定位网络定位缺陷集中分布的焊缝区域,再使用滑窗裁剪操作,将切割的小图输入到缺陷检测网络中进行精准的缺陷检测。然后,为了应对缺陷样本过少导致的网络过拟合问题,本文提出了基于负样本正常图像的Copy-Pasting数据增强策略,提升了缺陷检测精度。实验结果表明,本文提出的方法能够有效降低网络模型大小,同时mAP0.5达到99.5%,精确率99.8%,召回率99.6%,检测速度在20 frame/s至33 frame/s,能够满足实时辅助评片的要求。

X-ray nondestructive testing is an important method for weld quality inspection.In order to improve the efficiency and accuracy of X-ray image review.First of all,this paper proposes a real-time X-ray weld defect detection method based on lightweight YOLO network.On the basis of analyzing the characteristics of the weld radiographic inspection image,in order to improve the accuracy and speed of weld image defect detection,a cascade defect detection model is designed.First,the lightweight designed weld positioning network is used to locate the weld area where defects are concentrated,and then the sliding window cutting operation is used to input the cut small image into the defect detection network for accurate defect detection.And then in order to deal with the network over fitting problem caused by too few defect samples,this paper proposes a copy casting data enhancement strategy based on negative sample normal images,which improves the defect detection accuracy.The experimental results show that the proposed method can effectively reduce the size of the network model,and the mAP0.5 reaches 99.5%,the accuracy rate 99.8%,the recall rate 99.6%,and the detection speed is 20 frame/s to 33 frame/s,which can meet the requirements of real-time and aided review.