Research Of One-class Classification Strip Steel Surface Defect Detection Method Based On Generative Adversarial Networks

The machine vision-based strip steel surface defect detection system plays an important role in improving product quality and enhancing the intelligent manufacturing level.However,in actual production,it is difficult to obtain a complete and uniformly distributed defect sample database,and the machine learning method is not suitable because it relies on large amounts of labeled data.In anomaly detection,the one-class classification method can divide the normal samples and anomal samples by fitting the distribution of a large amount of normal data,and the generative adversarial networks(GAN)can build models for complex and high-dimensional data.Therefore,based on the one-class classification framework,the thesis proposes the one-class classification strip steel surface defect detection method based on the prior distribution guide generative adversarial networks(PDGGAN)and few defect simple guide generative adversarial networks(FDSGGAN),the contributions are as follows:(1)Although GAN can build models for complex and high-dimensional distributions of data,its application in extracting sample features and performing defect detection still requires further optimization of GAN.The prior distribution guide generative adversarial networks was proposed to automatically extract the features of the defect-free samples and complete the strip steel surface defect detection.By introducing instance normalization preprocessing and optimizing the processing of the prior distribution guide terms in the generator loss function of the PDGGAN,the description of feature distribution of the PDGGAN model and the accuracy of the extracted features can be improved.In addition,in response to the instability problem of the original GAN during the image feature distribution learning and the classification model the establishment,the hyper-parameter term of the balanced model was added to the discriminator’s loss function to improve the stability of the model.The experimental results show that,for the one-class classification framework,the model proposed in the thesis can more effectively learn image distribution features,and obtain more accurate defect detection results in comparison with traditional manual feature extraction methods and general GAN networks for defect detection.(2)In order to further suppress the interference of random textures for difficult-to-separate defect samples similar to the background texture,the one-class classification strip steel surface defect detection method based on FDSGGAN is proposed.Firstly,by adding a few defective samples in the discriminator model training process,the model can learn the similarities and differences of the defective samples and the defect-free samples to obtain better sample feature discrimination ability.In addition,the generator and discriminator models are balanced by adding a pull-away term(PT)to the loss function of the generator’s defective sample,and introducing an image diversity ratio factor into the discriminator loss function to further improve the stability of the model.Finally,the L1 norm is added to the generator loss function to sparse feature,thereby reducing the interference of the output-unrelated features on the detection result.The experimental results show that the classification defect detection effect based on FDSGGAN can obtain similar effects to the supervised strip steel defect detection method.