A Weld Defect Detection Method Based on Triplet Deep Neural Network
作者单位：School of Information Science and Engineering Northeastern University
会议名称：《第32届中国控制与决策会议》
会议日期：2020年
学科分类：080503[工学-材料加工工程] 08[工学] 0805[工学-材料科学与工程（可授工学、理学学位）] 0802[工学-机械工程] 080201[工学-机械制造及其自动化]
关 键 词：defect detection x-ray triplet loss deep neural network
摘 要：In industrial fields, Nondestructive Testing（NDT） has become an important method to test the quality of welds. For the low-contrast pipe weld defect x-ray image, the traditional detection method has low precision. In this paper, an automatic detection method for weld defects based on a triplet deep neural network is proposed. First, the original X-ray image is changed into a relief image, so that the feature of the defects is more obvious. Second, the feature vector is obtained by mapping the relief image through the triplet deep neural network. The deep neural network based on triplet makes the similar defect feature vectors are closer, and the distances of different defect feature vectors are farther. It is first time that the deep neural network based on triplet was used to detect the weld defect images. Finally, the weld defect was detected by Support Vector Machine（SVM） classifier. It is shown that the proposed detection method of weld defects has better performance than the conventional methods.