FeatDAE: Introducing Features With Denoising Autoencoder for Anomaly Detection

Anomaly detection (AD) is a critical task in manufacturing inspection. Reconstructive AD methods restore the normal appearance of an object, ideally modifying only the anomalous regions. However, previously commonly used reconstruction-based architecture always struggles with overgeneralization and overfitting problems, leading to poor reconstruction performance on real defective samples. In this study, we propose a more general denoising autoencoder, by introducing a feature hierarchy design to address these challenges in unsupervised AD. In particular, we operate feature transformation in the latent space to cope with the robustness of unseen anomalies in reality. Furthermore, the method enhances the discriminative capability of the model by focusing on multiple knowledge, including pixel color, histogram of oriented gradient (HOG) feature, and deep features. Additionally, a feature alignment module is proposed to manage the varied sizes and morphologies of features. Experiments conducted on both the MVTec AD dataset and the VisA dataset demonstrate that our FeatDAE significantly outperforms existing methods, achieving state-of-the-art results with high efficiency.