Unsupervised anomaly detection and localization with one model for all category

As most industrial products are defect-free, unsupervised anomaly detection and localization have become the focus of many researchers. In recent years, one-category-one-model algorithms have shown excellent performance on many datasets. However, algorithms in this paradigm are difficult and costly to maintain. In addition, existing algorithms that handle N categories with one model require a large number of samples to train the model, and their accuracy is low. To this end, we propose an unsupervised anomaly detection and localization algorithm with One Model for All Categories, referred to as OMAC. This method solves these problems by Lightweight Feature Extractors(LFE), Representativeness-based Sample Selection(RSS), and building Dual Memory Banks(DMB). We introduce the LFE to extract patch features and global features to reduce the time cost of model training and inference. To reduce the need for a large number of samples in existing methods, we propose an RSS algorithm to select representative samples for training the model. We propose a DMB algorithm based on a query mechanism to implement one model to detect all categories of products. Extensive experiments show that OMAC outperforms other state-of-the-art algorithms. Moreover, OMAC can achieve high frame rates of up to 58 FPS on the 3090 GPU, meeting the requirements of real-world factories.