Research On Key Technologies For Automatic Construction Of Semantic Segmentation Model For X-Ray Stainless Steel Weld Defects

The requirements for product quality in today’s society are constantly increasing with the development of industrial production technology.Due to a series of defects produced during stainless steel welding,the product can not meet the requirements of use.Therefore,it is particularly important to identify stainless steel welding defects,rapid and accurate identification of stainless steel weld defects has become a practical problem to be solved.At present,although some progress has been made in using deep learning method to detect weld defects,the following problems still exist: 1)Small weld defects are easily ignored by the neural network in the initial detection stage;2)As the convolution process deepens,the risk of invalid or even erroneous feature information being passed into advanced semantic feature processing modules increases;3)Most of the existing methods still remain on the basis of the identification of defect types,and cannot intuitively display the specific defect types and size;4)Most of the Non Destructive Testing(NDT)methods using deep learning at home and abroad need to design network models manually,which is not only time-consuming and labor-intensive,but also has poor generalization ability.In view of the above problems and difficulties,the main research carried out in this paper is as follows:(1)Research on the construction method of significant region enhancement and error correction attention module.Due to the insufficient attention paid by neural network models to subtle weld defects in the initial stage of feature extraction,it may lead to the occurrence of missed and misjudged weld defects.At the same time,existing semantic segmentation models inevitably generate invalid feature information when identifying defects in X-ray stainless steel welds,which is transmitted to advanced semantic feature processing modules.Therefore,this paper conducts research on the construction method of significant region enhancement and error correction attention module.Among them,saliency region enhancement module can be used to detect saliency targets in input images and generate the potential importance of each pixel,providing a reliable data foundation for subsequent feature extraction,and enhancing the attention of neural networks to saliency targets;The error correction attention module can be used to reduce the potential risk of information loss caused by incorrect judgments.(2)Research on module contribution theory for automatic construction of semantic segmentation model.Aiming at the problem that most of the existing semantic segmentation models are designed manually and are greatly influenced by subjective experience,and there is no quantitative visualization theory to support the automatic construction of semantic segmentation models.A module contribution theory for automatic construction of semantic segmentation model is proposed in order to accurately and efficiently calculate the module contribution to provide a new idea for automatic construction of semantic segmentation model,while reducing the research and time cost of artificial design network model.(3)Research on dynamic network programming algorithm based on module contribution theory.In order to further solve the problem that the results generated according to the module contribution theory still need to manually select candidate operations,taking into account the requirements of network model complexity and recognition accuracy under various application scenarios,a dynamic network programming algorithm based on the module contribution theory is proposed,so as to dynamically generate the network model with the best recognition accuracy under different constraints.Based on the above research contents,a semantic segmentation system for X-ray stainless steel weld defects was designed,which realized the semantic segmentation architecture search,dynamic network programming,single semantic segmentation,batch semantic segmentation and other functions,meeting the requirements of detection of weld defect size,category and location.Experiments show that the m Io U index of the proposed method reaches 49.12%in the semantic segmentation data set of X-ray weld defects,which is higher than that of all the semantic segmentation models in recent three years.The research in this paper is helpful to broaden the thinking of automatic detection of X-ray stainless steel weld defects,and plays a good role in complementing the existing in deep learning NDT technology,which plays a very important role in reducing the economic losses caused by the welding quality,improving the welding level,and reducing the safety liability accidents in the production process.