Automatic Detection of Weld Defects in Pressure Vessel X-Ray Image Based on CNN

作     者：XIAO Wenkai FENG Xiang NAN Shuiyu ZHANG Linlin XIAO Wenkai;FENG Xiang;NAN Shuiyu;ZHANG Linlin

作者机构：Department of Computer Science and Engineering.East China University of Science and TechnologyShanghai 200237China Shanghai Engineering Research Center of Smart EnergyShanghai 201103China Shanghai Aino Industrial Technology CO..LTD.Shanghai 201612China 

出 版 物：《Wuhan University Journal of Natural Sciences》 (武汉大学学报（自然科学英文版）)

年 卷 期：2022年第27卷第6期

页      面：489-498页

核心收录：

学科分类：08[工学] 081104[工学-模式识别与智能系统] 0811[工学-控制科学与工程] 

主　　题：nondestructive testing depth learning weld defect detection convolutional neural networks dilated convolution 

摘      要：The visual automatic detection method based on artificial intelligence has attracted more and more attention. In order to improve the performance of weld nondestructive defect detection,we propose DRepDet(Dilated RepPoints Detector). First, we analyze the weld defect dataset in detail and summarize the distribution characteristics of weld defect data, that is, the defect scale is very different and the aspect ratio distribution range is large. Second, according to the distribution characteristics of defect data, we design DResBlock module, and introduce dilated convolution with different dilated rates in the process of feature extraction to expand the receptive field and improve the detection performance of large-scale defects. Based on DResBlock and anchor-free detection framework RepPoints, we design DRepDet. Extensive experiments show that our proposed detector can detect 7 types of defects. When using combined dilated rate convolution network in detection, the AP50 and Recall50 of big defects are improved by 3.1% and 3.3% respectively, while the performance of small defects is not affected, almost the same or slightly improved. The final performance of the whole network is improved a large margin,with 6% AP50 and 4.2% Recall50 compared with Cascade RCNN and 1.4% AP50 and 2.9% Recall50 compared with RepPoints.