Research On Defect Detection Method Of X-ray Image Of Weld Based On Depth Learning

Welding defect detection is an important work in the welding industry,and the use of digital X-ray photography is an important means of welding non-destructive testing.Nowadays,manual methods are mainly used to identify defects in weld images,and there are many subjective factors in manual evaluation.In this paper,the automatic identification method of weld defects is studied deeply.In this paper,the real digital X-ray weld images are taken as the research object.In view of the large background area and many interference factors in the X-ray image,firstly,the weld in the X-ray image is extracted accurately.The corresponding image processing means are used to make the weld area stand out in the whole weld image.Threshold segmentation is carried out on the highlighted image of the weld,and the weld is segmented as the foreground.The edge of the weld area is optimized by morphological operation.The different foreground in the image is separated and the outer boundary is extracted,their key parameter values are obtained,and the weld region is screened out according to the key parameter values.The inclined weld is corrected,and the corrected weld area is framed with the minimum rectangle,and the precise extraction of the weld region is realized according to the coordinates of the frame.Carry on the defect annotation to the extracted weld image,and construct the weld data set.According to the characteristics of the data set,the UNet which has excellent representation on the small data set is used as the network infrastructure,the deep separable convolution is used in the under-sampling to reduce the number of parameters in the network,and the attention mechanism structure is used to suppress the redundant information.Hole convolution is used in the highest layer of under-sampling to improve the receptive field of high-level features and prevent the loss of edge details.In the upsampling stage,the multi-scale fusion strategy is used to obtain rich information under different sensory fields.In the training process,the joint loss function is used for training,which effectively improves the convergence speed of the model.Transfer Learning is used to improve training speed and detection accuracy.Finally,through experiments on the weld data set,the improved UNet model is objectively compared with the original UNet model,Seg Net model,FCN model and other similar methods,and the feasibility of the experimental results is analyzed.The experimental results show that the improved UNet model can be better applied to the defect segmentation of weld images.The defects segmented from the weld image are separated,each defect is located and the key parameters are obtained,and the defects and masks extracted by the positioning frame are used to realize the accurate extraction of the defects.classify and label the extracted defects.In order to solve the problem of data imbalance,SMOTE algorithm is used to balance the data.In the classification network model,a variety of network fusion methods are used as the feature extraction part,the SVD algorithm is used to screen the key features,and the attention mechanism module is added to the classifier to send the selected features to the classifier for classification.Experiments on the improved model are carried out on the weld defect data set,and the improved model is objectively compared with similar methods such as Dense Net model,Efficient Net model,Res Net model and so on.The experimental results show that the improved model achieves higher classification accuracy in weld defect classification.