A novel semi-supervised strip steel surface defect classification model based on CycleGAN and ResNet-50

The reliability and service life of steel products are largely dependent on the quality of the steel strip, so it is crucial to accurately identify its surface defects. Steel strip surface defect data collected from actual industrial scenarios is often limited and mostly unlabeled, so researchers often use Generative Adversarial Networks (GANs) for data augmentation while introducing semi-supervised learning to improve the performance of deep learning-based steel strip surface defect classification models. However, the samples generated by GANs often lack sufficient diversity and fail to simulate the actual distribution of the real defect data. In addition, strip surface defect data often exhibits highly similar texture features at the fine-grained level, which significantly degrades the performance of models. Therefore, a new semi-supervised strip steel surface defect classification model is proposed. Firstly, a data augmentation strategy based on CycleGAN is developed to generate unlabeled samples with more diversity and fit real industrial scenarios, thereby improving the model's generalization ability. Secondly, a feature extraction and classification network based on ResNet-50 is designed to enhance the model's capability to capture fine-grained features, thus improving the ability to distinguish between different defect categories. Finally, extensive experiments were conducted on the NEU-CLS dataset. The results showed that when trained with the 10% labeled data and the 90% unlabeled data, the proposed model achieved accuracy, precision, and recall rates of 92.91%, 89.64%, and 84.81% in this dataset, indicating competitive performance compared with multiple existing semi-supervised models.