GLASNet: A Global–Local Anomaly Generator and Salient Channel Feature Selector-Based Network for Accurate Few-Shot Industrial Anomaly Detection

Anomaly detection of products is of great significance in modern industrial production. However, the current mainstream unsupervised anomaly detection methods based on deep learning are still challenging in the face of the cold start of cross-product migration in real industrial production. Here, an innovative few-shot anomaly detection (FSAD) model, global–local anomaly and salient selection network (GLASNet), comprising a global–local anomaly generator, a salient channel feature selector, and a phased score mapping loss is developed. Reliable global and local anomaly features based on a Gaussian distribution at the feature level can be generated through the elaborately designed global–local anomaly generator, simulating a more comprehensive generation of anomaly samples. This helps the model to better learn the feature distribution of anomaly samples, guiding it to focus more on local subtle anomalies and alleviating data imbalance problems in few-shot scenarios. Additionally, the positive and negative features are mapped to a new feature space at multiple scales by the salient channel feature selector to eliminate feature domain bias. Representative features are then selected in the channel dimension to reduce the interference of information redundancy on feature mapping. Finally, combined with a novel phased score mapping strategy, multistage guidance of feature mapping is carried out. As a result, clear positive and negative feature boundary contours and uniform distribution of similar features can be obtained, enabling the model to make more accurate discrimination of features with strong similarity. The results of multiple experiments verify the effectiveness of the proposed modules and GLASNet outperforms the state-of-the-art FSAD methods in terms of image-level AUROC, achieving improvements of 0.5%–2.9% on the MVTec AD and MPDD benchmarks.