Research On Weld X-ray Image Defects Recognition Based On Deep Learning

In the welding production process,affected by various welding parameters and the external environment,it is easy to form various welding defects of different degrees and numbers in the weld.However,the artificial X-ray image defect detection and recognition method is need to improve in both efficient and accurate.Based on the Xray images of pipe welds,this paper studies the automatic detection and recognition methods of weld defects.The purpose is to use deep learning methods to achieve better identification of weld defect identification,thereby improving the accuracy of weld defect identification and enabling It is more efficient,standardized and intelligent.During the defect detection process,for the X-ray image of the pipe weld,median filtering technology is used to remove noise,and image enhancement technology is used to increase the contrast of different areas in the weld to improve the influence of different areas contrast in subsequent image processing.the weld area was segmented by the Maximum Between-Class Variance method,and the defect edge of the weld was detected by Sobel edge detection technology.Aiming at the problem of image area estimation in defect feature engineering,chain code tracking and pixel statistical methods were used to effectively solve the area estimation of circular defects,and the cross-coordinate positioning method was used to locate them.The normal images of pores,cracks,non-fusion,non-penetration defects and non-defects are extracted and data enhancement and size normalization operations are performed to complete the preprocessing of the weld image to construct a sample image data set.In the defect recognition process,deep learning technology is used to identify defects in the weld image.Firstly,some problems in the conventional convolutional neural network model are analyzed,and an improved activation function and an adaptive pooling method considering the pooling domain are adopted in the model,which is applied to the recognition task of weld defect images.By constructing a new convolutional neural network model and training the CNN model with the sample image data set as an input sample,the comparative experimental test shows that the designed CNN model not only has good robustness,can accurately identify image defects,and eliminating the complexity of manually extracting features,ultimately achieving a high recognition accuracy of 98.13%,thereby improving the degree of automation of detection and recognition.Aiming at the problem of deep neural network model recognition on weld inspection image dataset,a transfer learning method based on convolutional neural network was proposed.With the help of pre-trained models on large data sets,the pretrained models in the source data domain are transferred to the weld inspection image data set based on the difference between the content in the source data domain and the target data domain.By using the frozen layer method to train the parameters of different layers and fine-tune the network,the effectiveness of transfer learning in defect identification of weld inspection images was verified,and the effect of freezing different layers on the recognition performance of the model was studied.