Unsupervised surface defect detection of aluminum sheets with combined bright-field and dark-field illumination.

• An unsupervised surface defect detection of aluminum sheets is proposed. • A patch-aware pseudo-siamese feature matching network is developed. • A scoring function considering multiple aspects among the samples is designed. • A dataset of defect samples captured from aluminum production lines is constructed. In the field of quality control for aluminum sheets, the presence of surface defects can greatly impact the overall product quality. Supervised machine learning methods, which necessitate a substantial quantity of expertly annotated defect samples, present limitations when faced with a limited number and lack of diversity of defect samples. To address these challenges, an unsupervised defect detection method, which utilizes a combination of bright-field and dark-field illumination, is proposed. Combined bright-field and dark-field illumination can provide more information about the surface and enhance the visibility of defects by illuminating the surface from different angles and at different levels of reflectivity. Based on the combined illumination condition, an unsupervised patch-aware feature matching network, which is inspired by the anomaly detection paradigm and only requires defect-free samples for training, is proposed in this paper. The network can extract high-resolution features from bright-field and their corresponding dark-field images simultaneously. Additionally, a well-designed scoring function that considers both intra-field and inter-field relationships is introduced to obtain more accurate anomaly scores for the score map. Moreover, artificially simulated abnormal samples are incorporated into the training phase, which assists the network in explicitly learning potential differences between normal and abnormal samples. The proposed method was thoroughly evaluated on a dataset of surface defects in aluminum sheets. The experimental results demonstrate the superior performance of the proposed method in defect identification and segmentation, achieving a high level of accuracy with small model size and short inference time, outperforming other neural network-based methods. The method has been implemented in a real-time machine vision system, resulting in a significant improvement in detection efficiency and product quality. [ABSTRACT FROM AUTHOR]