Surface Defect Detection Based on ResNet Classification Network with GAN Optimized

Surface defect detection plays an important role in industrial domains. The defect failed to report and the defect missed to report will have a serious impact on the whole industrial production process, such as production stagnation. Traditional defect detection methods are based on image processing or shallow machine learning techniques, but these can only detect defects under specific inspection conditions. And the related methods of artificial intelligence still have some flaws and challenges, including small data sets, poor model accuracy, over-fitting generalization, and others. In this study, we offer a novel method for surface defect detection based on GAN optimized ResNet classification training on data sets, which creatively integrates the two networks. In specifically, the ResNet network training model is used together with the generative adversarial network (GAN) for defect detection and classification. The GAN is used to first supplement the original data, which is limited in quantity and category. Then the data and two models are organically integrated. We find that this GAN-optimized strategy outperforms direct classification using the ResNet network alone after conducting several 200-epoch runs on a variety of big defect pictures.