| Welding technology has been widely used in a variety of key areas such as aerospace,defense science,machinery manufacturing,and petrochemical.X-ray inspection is the main means of weld defect detection.Manual identification has problems such as false detection,missed detection,and inefficiency due to subjective factors.Weld defect detection methods based on deep learning can greatly improve detection efficiency and accuracy.Deep learning techniques require a lot of data,but the lack of open data sets in the field of weld defect detection has hampered further research.This paper expounds the relevant research status at home and abroad for the topic of weld defect detection based on deep learning.By studying the architecture of the dataset,the characteristics of the image dataset and the characteristics of weld defect detection,the organization form of the weld image dataset are proposed,and a new weld image dataset WDXI is constructed and released.The image preprocessing process was developed according to the characteristics of the weld image,including grayscale transformation,image noise reduction and weld area cropping.In the defect detection part,the method of detecting the weld defect based on the sliding window is first proposed.The image in the window is classified by the WNet,and finally generated six probability maps.The Selective Search algorithm and non-maximum suppression are used to generate bounding boxes from the probability map.Secondly,the target detection framework Faster-RCNN which is widely used in the deep learning field is used for defect detection.The network structure of Faster-RCNN is studied in detail.The problem of inaccurate detection of small-scale targets is optimized first,and then the accuracy of detection is improved by using ResNet instead of VGGNet.In this paper,after the original weld image is sorted,labeled and image preprocessed,the WDXI data set is constructed and provided to the researchers.The data set will become a public data platform to complete the method verification and result comparison.The automatic detection method based on Faster-RCNN is better than the sliding window method.The accuracy of position and type is 62.40%.The results verify the feasibility and effectiveness of the method.The detection method based on Faster-RCNN combines the discrimination of defect type and location,which opens up a new direction for future research. |
