Context-aware feature reconstruction for class-incremental anomaly detection and localization.

With the development of deep learning, the unsupervised visual anomaly detection and localization task has gained significant attention in both academia and industry, where only normal data are used for training. Existing methods for this task typically train using all training data simultaneously. However, in practical industrial scenarios, new product classes are usually introduced incrementally, leading to the sequential availability of training data. Such scenarios demand methods for class-incremental anomaly detection and localization (CADL). The main challenge of class-incremental learning is to retain knowledge of old classes when learning new classes. In this article, we aim to effectively leverage limited exemplars of old classes to retain knowledge for the CADL task. Achieving this goal requires a model that can efficiently capture anomaly-identification-related knowledge from limited exemplars. Considering that pixel-level anomaly identification requires an understanding of the surrounding context, we treat context within inputs as valuable anomaly-identification-related knowledge and design a context-aware feature reconstruction (CFR) model to capture such knowledge. Moreover, to avoid inter-class context conflict that may arise with class increments, we design an intermediate feature organization strategy. This strategy and output-level knowledge distillation jointly form dual constraints to regularize the model at both mid-feature and output levels. Utilizing the CFR model with Dual Constraints, the proposed CFRDC can effectively retain old-class knowledge while learning new classes, thus addressing the CADL task. Experimental results on the commonly-used MVTec-AD dataset demonstrate the effectiveness and outstanding performance of the proposed method in CADL.