Multi-defect detection for magnetic tile based on SE-U-Net

Vision-based on-line full detection of surface defects is of great significance to the production efficiency and quality of magnetic tiles. In this work, we introduce a pixel-wise surface defect detection model called SE-U-Net, which simultaneously realizes detection and classification. The image augmentation based on instance transfer effectively reduces the imbalance between the background and the defect. The squeeze-and-excitation module empowers U-Net to adaptively learn the shallow information matched with deep features in the skip path, enhancing the ability to identify small defects with a few of additional network parameters. Experiments show that the recognition accuracy of this model exceeds the existing methods, and the mean pixel-accuracy reaches 0.97.