Research On Identification Of Multi-Type Welding Defects In X-Ray Images Based On Deep Convolutional Neural Network

The use environment of modern industrial weldments has higher requirements on the welding quality of the weldments.However,due to many uncertain factors in the welding process,the welding part is prone to produce welding defects such as cracks,holes,slag inclusion and incomplete fusion/penetration.Welding defects will seriously affect the quality and performance of weldments.In order to prevent engineering accidents,it is of great significance to carry out efficient welding defect identification.X-ray images are often used in the detection and recognition of welding defects inside weldments,but the artificial welding defect recognition technology based on X-ray image is limited by the technical level of inspectors and other factors,the recognition efficiency is low,and the recognition results have strong subjectivity.With the development of computer technology,the X-ray image welding defect recognition technology based on convolutional neural network has gradually emerged.However,at this stage,the welding defect detection models established based on deep convolutional neural network have not been improved according to the characteristics of welding defects,so the overall detection accuracy of the welding defect detection model is low.In order to improve the overall detection accuracy of the welding defect detection model,this paper improves the existing welding defect detection model.The specific research contents are as follows.On the general target detection data set,the Retina Net model directly uses the preset size and scale to generate the anchor frame,but this method does not work well on the small-size welding defect data set.To solve this problem,an adaptive anchor frame generation method for Retina Net welding defect detection model is proposed in this paper.According to the characteristics of the welding defect data set,the method adaptively generates the detection anchor frame,reduce low quality training samples during training,and avoids the defect features of the model learning errors,so as to improve the overall detection accuracy of the Retina Net welding defect detection model with the input of low-resolution images.In the Retina Net welding defect detection model,the locality of the residual learning unit structure limits the feature extraction capability of the network,the method of generating positive and negative samples with fixed intersection ratio introduces low-quality positive samples.For these problems,this paper proposes an improved Retina Net model.The model uses an adaptive training sample selection algorithm to improve the quality of training positive samples;and adds the Res2 Net module to the feature extraction network to enable the model to learn richer semantic information of welding defects;finally,the attention module is used to improve the feature pyramid network so that the network model pays more attention to the valid defect features.The experimental results show that compared with the original Retina Net welding defect detection model,the improved Retina Net’s multi-type welding defect detection model can improve the overall detection accuracy and recall by 9.4% and 2.9% respectively.The overall detection accuracy of the improved multi-type welding defect detection model is higher than the existing mature target detection model,which verifies the effectiveness and advancement of the improved Retina Net multi-type welding defect detection model.