Abstract

Industrial image anomaly detection is crucial for identifying abnormal patterns and their locations in manufacturing. Memory bank methods have recently gained attention for their high performance without requiring additional training. However, these methods often struggle to capture the unique characteristics of industrial training data. This is because feature extractors pre-trained on general natural images may produce features that do not align with the specific properties of industrial images, leading to a mismatch with learning objectives. This discrepancy can cause memory bank clusters to inaccurately represent normal data by either over-generalization or over-specialization. To overcome these challenges and improve upon existing memory bank based approaches, we propose a method that effectively learns the characteristics of the training data and optimizes the representation space to improve the differentiation between normal and abnormal patterns. Our method uses a learned projection layer to project patch embeddings into a new representation space, forming clusters that clearly represent the unique characteristics of training data. Our method achieves this by using coreset-based pseudo-labels and proxy metric learning. The projection layer is trained to bring similar patch embeddings closer and dissimilar ones farther apart, with each pseudo-label as the center. Experiments on various benchmark datasets demonstrate that our method outperforms existing state-of-the-art methods. Furthermore, by effectively adapting to the specific domain of the training data, our proposed methodology enables precise anomaly localization. This work contributes a novel implemented artificial intelligence solution to the challenging application of artificial intelligence in industrial inspection, demonstrating superior performance and precise localization. The code is available at : https://github.com/crimama/ProxyCore.