Hybrid Encoder for Anomaly Detection Based on Latent Feature Regularization

Anomaly detection aims to enable the model to accurately distinguish anomalies that deviate from normal data. Unsupervised learning relies on generative models to reconstruct normal data, assuming the model learns feature distribution of normal data, treating the anomalies as boundaries away from normal data. However, real-world anomalies are small, morphologically variable, challenging to predict accurately, and closely resemble normal data in both appearance and la-tent space distribution. To tackle this issue, this paper intro-duces a unique masked mechanism and a straightforward noise injection approach. By introducing noise into the feature space of normal samples and subsequently employing a noise-reducing decoder for denoising, the normal data becomes more condensed in the latent space, facilitating the detection of anomalous outliers. Our approach achieves a 76.3% average anomaly localization accuracy on the MVTec AD dataset and effectively detects small-area anomalies without depending on methods such as ImageNet pre-trained models.