Research On Defect Recognition And Localization For X-Ray Welding Images Based On Convolutional Neural Network

In the welding process,it is inevitable to produce defects,which mainly include blow hole,incomplete penetration,crack and so on.These defects cause huge safety risks and potential property losses to our industrial production and daily life.X-ray technology is often used to project the weld area to produce a film image of the weld to evaluate the welding quality of the weld,so as to determine the type and location of defects.At present,this judgment process mainly relies on manual work,but it has many disadvantages such as low efficiency and strong subjectivity.In the field of deep learning,Convolutional Neural Network(CNN)has been extensively studied and applied in various application fields of images,and has achieved remarkable results.Therefore,the identification and location of weld defects can be accomplished well by computer vision technology.Aiming at the above problems,this paper designs a defect recognition and location method based on deep learning for X-ray welding image.Firstly,a lightweight feature extraction module RG-CSP is designed.This module can reduce the parameters and computations while ensuring effective information learning,and achieve the balance between performance and lightness as far as possible.Based on the RG-CSP(Residual Ghost-Cross Stage Partial)module,this paper proposes a novel LF-YOLO(Lighter and faster-YOLO)network for defect recognition and location.This network is mainly constituted by the RG-CSP module,and has a shallow network depth.These characteristics make the LF-YOLO model have a lower complexity,which can achieve faster running speed and lighter model size.This paper also proves the information extraction ability of LF-YOLO for weld defects by means of feature maps visualization.Finally,through a large number of experiments,the ability of LF-YOLO to identify and locate defects is proved.LF-YOLO achieves 88.0 m AP(mean Average Precision)with111 FPS(Frame Per Second)prediction speed and 20.3 MB model size,which verifies that it can greatly reduce the resource requirements for carrying equipment,and prove the practicality and economy of LF-YOLO.