Deep Embedded Auto-encoder for End-to-End Unsupervised Image Anomaly Detection

Abstract

Image anomaly detection plays a critical role in industrial quality control, medical diagnostics, and security surveillance, yet existing unsupervised methods often suffer from limited detection accuracy and poor adaptability. To overcome these limitations, we propose UAD-ADC, a novel framework that automatically identifies anomalies in images without requiring labeled training data. Our approach integrates deep representation learning with adaptive clustering to effectively separate normal patterns from anomalies by learning robust feature representations and dynamically refining decision boundaries. A key innovation is our intelligent sample selection mechanism, which enhances model stability by prioritizing high-confidence normal samples during training, along with an iterative optimization strategy that progressively improves anomaly discrimination. Extensive experiments on benchmark datasets demonstrate that UAD-ADC significantly outperforms state-of-the-art unsupervised methods, with particular effectiveness under varying anomaly ratios. These advancements pave the way for more reliable and scalable unsupervised anomaly detection in practical scenarios.