MFFA: Multi-level feature fusion and anomaly map compensation for anomaly detection

Embedding similarity-based methods obtained good results in unsupervised anomaly detection (AD). This kind of method usually used feature vectors from a model pre-trained by ImageNet to calculate scores by measuring the similarity between test samples and normal samples. Ultimately, anomalous regions are localized based on the scores obtained. However, this strategy may lead to a lack of sufficient adaptability of the extracted features to the detection of anomalous patterns for industrial anomaly detection tasks. To alleviate this problem, we design a novel anomaly detection framework, MFFA, which includes a pseudo sample generation (PSG) block, a local-global feature fusion perception (LGFFP) block and an anomaly map compensation (AMC) block. The PSG block can make the pre-trained model more suitable for real-world anomaly detection tasks by combining the CutPaste augmentation. The LGFFP block aggregates shallow and deep features on different patches and inputs them to CaiT (Class-attention in image Transformers) to guide self-attention, effectively interacting local and global information between different patches, and the AMC block can compensate each other for the two anomaly maps generated by the nearest neighbor search and multivariate Gaussian fitting, improving the accuracy of anomaly detection and localization. In experiments, MVTec AD dataset is used to verify the generalization ability of the proposed method in various real-world applications. It achieves over 99.1% AUROCs in detection and 98.4% AUROCs in localization, respectively.