Research progress of metal surface defect detection based on deep learning

Detecting metal surface defects is essential in industrial production to ensure product quality and safety. This study reviews the evolution of defect detection techniques, starting from traditional methods and progressing to deep learning-based techniques, particularly the use of Convolutional Neural Networks (CNNs) and their advanced variants like ResNet, DenseNet, and Inception. In addition, transfer learning and data enhancement techniques can solve the problem of low accuracy of models with fewer data samples. In recent years, Graph Embedding and Distribution Transformation (GEDT) methods, Feature Graph Reconstruction Networks (FRNs), Cascade R-CNN, and CA-AutoAssign have continuously improved defect detection accuracy. In this paper, the FSC metal surface defect dataset is used, and the CA-AutoAssign method utilizing the CSPDarknet backbone achieves the highest recognition accuracy of 82.36% on the FSC dataset, outperforming other models. Despite significant progress in deep learning-based methods for detecting defects on metal surfaces, many challenges remain. A key focus should be improving the models' ability to generalize across different domains, particularly for use in varied and complex industrial environments. Future research should investigate adaptive techniques and expand the range of training datasets to enhance model performance in more challenging scenarios. Further analysis is needed to understand the factors that influence model performance under different conditions, which would help optimize defect detection methods for practical industrial applications.