Research On X-ray Weld Defect Detection Based On Improved FCOS

Welding is an important part of the industrial production process and most metal products are manufactured and processed by welding technology.However,in the actual production process,welded structures often produce a variety of defects such as porosity that affect the quality of the product.Therefore,it is particularly necessary to carry out quality checks on the weld seams of welded parts.One of the most common methods used in industry is radiographic inspection.The traditional method of X-ray weld defect detection relies on manual inspection of the film,which is inefficient and has a high leakage rate and does not meet the needs of modern production.Nowadays,computer vision technology is rapidly developing in the field of defect detection with its advantages of speed,accuracy and contactlessness.This thesis combines computer vision technology and deep learning technology to investigate X-ray weld defect detection,which is divided into the following main points.（1）Classify the X-ray weld defect images acquired in this thesis into six categories in accordance with the industrial standard（"Radiography of welded joints in metal fusion welding" GB/T3323-2005）,and analyse the morphological characteristics of each type of defect,then deploy an experimental platform for X-ray weld defect detection based on deep learning.（2）This paper pre-processes various types of weld defect images in a targeted manner,including image enhancement,image noise reduction and image expansion using generative adversarial networks,and finally evaluates the image quality by combining IS and FID indicators to complete the dataset.（3）Selecting various representative models of target detection algorithms,conducting comparison experiments on the data set of this thesis,analysing the advantages and disadvantages of each algorithm and combining them with the application scenario of this thesis,this thesis proposes an improved anchor-free detection algorithm FCOS-RFP.The algorithm uses a recursive pyramid as the feature extraction structure,and modifies the backbone network by adding feature branches to improve the feature extraction capability.In the prediction network part,the loss functions of the three branches are optimised to improve the flexibility of the model.The experimental results show that the FCOS-RFP algorithm can improve the accuracy of the two types of small-sized weld defects,namely porosity and slag,by 9.6% and 6.2% respectively,compared to the original algorithm,and the overall average accuracy of the model is improved by 4.8%.In summary,X-ray weld defect detection is regarded as the research topic of this thesis,and the target detection model designed based on deep learning effectively improves the detection accuracy of small-sized weld defects,providing its own innovative solution for progress in the field.