Research On Visual Inspection Method Of Metal Surface Defects Based On Generative Adversarial Networks

As one of the main raw materials in the field of industrial production,metal materials are widely used in military,petrochemical and other key industries.Due to the production process and manufacturing environment and other complex factors,its surface is easy to produce all kinds of defects.These defects not only affect the appearance of the product,but also adversely affect its performance and safety.Therefore,it is very important to detect metal surface defects to control their quality.Traditional machine vision defect detection methods have some shortcomings,such as unsatisfactory detection effect,high application environment requirements,low detection efficiency and poor adaptability.With the development of computer technology,methods based on deep learning are introduced into the field of metal surface defect detection.This dissertation mainly studies the detection method of metal surface defects based on Generative Adversarial Network,and improves the Generative Adversarial Network according to the characteristics of defect detection.The main research content is divided into the following three parts:1.Aiming at the problem of small samples due to the difficulty of obtaining a large number of labeled defect samples,a method for generation and classification of metal surface defects is proposed based on improved auxiliary classification generative adversarial network.Firstly,the residual blocks are used to optimize the basic network structure,strengthen the mining performance of the network for deeper features,and the spectral norm normalization is introduced into the convolution layer of the model to prevent the abnormal gradient change of the model and improve the stability of training.Secondly,the attention mechanism module is introduced into generator and discriminator to improve the ability of learning and generating defect features.At the same time,in order to alleviate the mode collapse problem of generative adversarial network,the gradient penalty is added to the loss function to constrain the sharp gradient change of the discriminator.Finally,the sample expansion is realized through adversarial optimization training.After training,the model is applied to the NEU strip surface defect dataset and the actual impeller machining surface defect dataset,and good results are obtained.2.To solve the problem that labeled defect data is difficult to obtain and a large number of unlabeled samples are not properly used in actual detection,a surface defect detection model based on improved semi-supervised auxiliary classification generation adversarial network is proposed.Firstly,based on the idea of positive-unlabeled learning,the loss function of the discriminator network is improved to improve the quality of generated samples and enhance the stability of the semi-supervised model.Secondly,in order to improve the accuracy of semisupervised defect classification,a threshold filter based on label confidence is established.While ensuring the classification accuracy of synthetic samples,the threshold filter filters the unlabeled data with high confidence pseudo labels to update the classification loss,which improved the utilization rate of unlabeled data.Finally,the validity of the model is verified on the NEU strip surface defect dataset and the actual impellers surface defect dataset.The model combined sample generation and semi-supervised learning to improve the accuracy of defect detection in the case of fewer labeled samples.3.Due to the difference in data fields and insufficient supervised samples in actual production environment,a new model for metal surface defect detection is established based on the idea of generative adversarial network and unsupervised domain adaptation.Firstly,in view of the problem that traditional machine learning is limited by the same distribution of data and cannot solve the difference of data distribution,the idea of domain adaptation is introduced to make full use of the feature information of source domain data to assist the defect detection task of target domain in the case of unlabeled;Secondly,in order to make the transferred information more accurate,the encoder and decoder networks based on adversarial training is used to learn the public features between domains to avoid the negative transfer phenomenon caused by the private feature transfer in each domain,and improve the adaptability of the model.At the same time,in order to avoid the over-fitting problem caused by insufficient migration information to represent the feature space,the adaptive feature embedding enhancement method is introduced to fully mine the feature information and improve the accuracy of defect detection.Finally,the experimental verification is carried out on the strip surface defect dataset and the actual industrial surface defect dataset.