Semi-Patchcore: A Novel Two-Staged Method for Semi-Supervised Anomaly Detection and Localization

Product surface defect detection is a crucial technology in industrial production. The adoption of deep learning-based algorithms for inspecting product surface defects has been steadily increasing due to their superior detection capability and enhanced generalization performance. However, current deep learning-based algorithms primarily focus on supervised approaches, which can be inefficient and costly. In this article, we present a novel pipeline for semi-supervised defect detection called Semi-Patchcore, which achieves comparable defect detection performance to weakly supervised methods using only defect-free and unlabeled samples for training. Initially, we establish a memory bank using a labeled defect-free training dataset. Subsequently, we compare the unlabeled mixed data with the features in the memory bank to derive pseudo-class labels. Finally, we train a segmentation network based on DeepLabV3+ using the pseudo-classification labels. To evaluate the performance of our approach, we conduct comparative experiments on four public datasets: MVTecAD dataset, DAGM dataset, BTAD dataset, and KSDD2 dataset. The experimental results demonstrate that our method outperforms state-of-the-art semi-supervised or unsupervised methods in terms of superiority and generalization. Additionally, we explore the impact of label issues on supervised learning observed in this study. Our method also surpasses some weakly supervised segmentation algorithms, showcasing its effectiveness in industrial defect detection.