Steel Surface Defect Detection Using GAN and One-Class Classifier

Automatic strip steel surface defect detection is a difficult mission, as a result of the imbalanced class distributions caused by the sparse distribution of abnormal samples. The one-class classification (OCC) method can detect abnormal samples by only training the normal samples. The Generative Adversarial Networks (GAN) can automatically learn the features of samples in unsupervised situations, and only one sample is used to train the model. The GAN-based one-class classification method for strip steel surface defects detection is proposed in the paper. The second to last output layer of GAN generator is chosen as the feature, which contains some basic and important information about the sample. In addition, an improved loss function is proposed to raise the stability of the model and the convergence speed. Then the one-class classifier can easily detect abnormal samples by comparing the feature of normal samples and abnormal samples. The proposed approach is validated in the strip steel data sets containing surface defect of different size, shape and type. The experiments have shown that the method can reach an average accuracy of 94% in the data sets.