Model-guided contrastive fine-tuning for industrial anomaly detection

State-of-the-art industrial visual anomaly detection now relies on modeling the distribution of pre-trained neural network features. To this goal, most of the work has focused on how to model features of normal data and the choice of the pre-trained network. The current trend is to use a network pre-trained using self-supervised contrastive learning so that the same network can be used for all possible downstream applications. However, this also means that the network is object and task agnostic, meaning that features are very generic and not optimized with the detection model. In this paper, we propose to look at how to specialize features for a given application so as to improve performance and propose a fine-tuning process taking advantage of the differentiability of some popular models. This fine-tuning is performed following a contrastive learning framework meaning that no real anomalies are necessary during the process. We demonstrate the improvement on both localization and quality of detection on the MVtec dataset.