Stable Latent Anchoring for Enhanced Anomaly Detection in Convolutional Autoencoders
TL;DR
A convolutional autoencoder (CAE) framework with a dual-term loss that combines reconstruction accuracy with a latent anchoring mechanism, encouraging latent representations to align toward a reference vector can achieve competitive performance compared to more complex models while maintaining simplicity and ease of deployment for practical anomaly detection applications.
Abstract
Autoencoders (AEs) are often limited by their reliance on pixel-level reconstruction errors, which restricts their ability to capture high-level semantic features for anomaly detection. To address this limitation, we propose a convolutional autoencoder (CAE) framework with a dual-term loss that combines reconstruction accuracy with a latent anchoring mechanism, encouraging latent representations to align toward a reference vector. This design promotes compact and structured latent spaces, enabling improved separation between normal and anomalous samples. In addition, the impact of image preprocessing techniques, including Gabor filtering and per-image standardization, is investigated. For anomaly scoring, both distance-based and density-based approaches are applied in the latent space using k-nearest neighbors (KNN) and Gaussian mixture models (GMM). The proposed method is evaluated on MNIST, Fashion-MNIST, and CIFAR-10 datasets. Experimental results demonstrate consistent improvements in mean area under the receiver operating characteristic curve (AUROC), increasing from 0.9688 to 0.9853 on MNIST, from 0.9365 to 0.9445 on Fashion-MNIST, and from 0.6016 to 0.6931 on CIFAR-10. These findings show that the proposed framework can achieve competitive performance compared to more complex models while maintaining simplicity and ease of deployment for practical anomaly detection applications.