摘要
目前基于深度学习的卷积神经网络在对焊缝缺陷射线图像进行智能识别时,多采用基于候选区域的two-stage方法,检测速度难以满足实时性要求。针对该问题,提出基于深度卷积神经网络的one-stage焊缝缺陷识别定位算法,将整张图像输入网络,并直接在输出图像上标定目标缺陷位置及类别。通过采用特征金字塔、减小网络深度、引入跳跃连接卷积块和K-means算法等方法对YOLO网络进行改进,提高了网络对焊缝缺陷识别定位的准确率和速度。实验结果表明:该方法比two-stage识别定位算法和YOLO原网络在检测速度和检测精度方面都有所提升,单个图像的平均识别准确率为94.9%,召回率为94.1%,处理时间为19.58 ms,具备焊缝缺陷在线实时识别的工程应用价值。
Currently,a two-stage convolution neural network object detection algorithm based on region-proposals is widely used to identify and locate defects in weld radio graphic images.However,its detection speed does not meet the real-time requirements.In order to solve this problem,a one-stage weld defect identification and location algorithm based on deep learning network is proposed in this paper.In this algorithm,we input the whole image into the network,and directly calibrate the target defect position and category on the output page.With the YOLO network improved by using the feature pyramid,reducing the depth of the network and introducing densely-connected-convolutional-block and K-means algorithm,its accuracy and speed to identify and locate weld defects are improved.The experimental results show that this proposed algorithm is superior to the two-stage recognition and localization algorithm based on candidate regions and the original network of YOLO in the aspects of detection speed and accuracy and that its average recognition accuracy rate of a single image is 94.9%,the recall rate is 94.1%,and the processing time is 19.58 ms.Thus,this proposed algorithm is of great value in the engineering application in on-line real-time identification and location of weld defects.
作者
李砚峰
刘翠荣
吴志生
孙前来
朱彦军
李科
LI Yan-feng;LIU Cui-rong;WU Zhi-sheng;SUN Qian-lai;ZHU Yan-jun;LI Ke(School of materials science and engineering,Taiyuan University of science and technology,Taiyuan,030024,China;School of Electronic Information Engineering,Taiyuan University of science and technology,Taiyuan,030024,China)
出处
《广西大学学报（自然科学版）》
CAS
北大核心
2021年第2期362-372,共11页
Journal of Guangxi University（Natural Science Edition）
基金
国家重点研发计划课题(2018YFA0707305)
山西省自然科学基金资助项目(201901D111265,201801D121082)。
关键词
深度学习
焊缝缺陷
智能识别
deep learning
weld defect
end-to-end identification and location
