Recognition Of X-ray Typical Defect Features Of Welds Based On Deep Learning

Welding technology is the most important method to connect various metal components,which has been widely used in crane,pressure vessel and other special equipment.The quality of welding process directly affects the service life of the welding structure.Due to technical limitations,some welding defects will inevitably occur in the welding process.At present,nondestructive testing is used to test defects.Among them,X-ray detection technology can directly and reliably reflect the shape,location and other information of defects,so it is widely used in the field of weld nondestructive testing.At present,the images obtained from X-ray images are basically manually examined,and the subjective factors such as professional level will lead to missed or false detection and low efficiency.The machine vision inspection technology developed in recent years also relies on manual selection of defect features,and sometimes it is impossible to select appropriate features.When the quality of the extracted defect features is not good,the accuracy of the algorithm cannot be guaranteed.This paper aims to establish the automatic identification model of typical weld defects through automatic learning characteristics of deep learning technology,and improve the accuracy,efficiency and automation level of detection.The second chapter expounds the theoretical basis of deep learning application in the field of defect recognition,and finally selects Faster R-CNN as the basic model of this topic through analysis and comparison.In the third chapter,aiming at the lack of image data,the data amplification methods such as horizontal flip,image rotation and image translation are used to amplify the DR images of welds,and 11985 data sets of four typical weld defects are established.Aiming at the problems of image noise and low contrast,the noise characteristics and defect characteristics are analyzed.The median filter is used to eliminate the noise in the image,and the power law(gamma)transform is used to improve the image display quality.Combined with the characteristics of weld defects in the fourth chapter,the Faster R-CNN is improved adaptively.The residual network Res2 Net is used to replace the original backbone network to improve the feature extraction ability,and the weighted feature fusion module is added to combine the rich semantic information of the high level and the rich contour texture information of the low level.The feature maps of each layer are predicted respectively,so as to improve the detection performance,especially for small targets such as circular defects.The experimental platform is built and the effectiveness of the improved model is verified.The detected defects are quantitatively analyzed according to national standards.Finally,a set of weld defect detection system is designed and developed.In the process of verifying the effectiveness of the improved model,ensuring the same hardware platform and experimental parameters,the mean average precision of RBFatser?rcnn model with weighted feature fusion and Res2 Net replacing original backbone network is 85.5 %,and the detection time of single DR image is 0.588 s,which is 4.2 % and 8.2 %higher than that of RFaster?rcnn model with Res2 Net replacing backbone network and original Faster R-CNN model.The test results show that the improved model avoids the complexity of artificial feature selection,and has the advantages of high defect recognition accuracy and fast detection speed,which has important practical value in the field of X-ray detection.