Unsupervised Image Anomaly Detection: A Review of Autoencoder and Improved Clustering Fusion Methods

Abstract

Unsupervised image anomaly detection requires no labeled data and holds broad application prospects in fields such as industrial quality inspection and medical imaging. The dual-driven approach combining autoencoders and clustering techniques represents one of the core research directions in this field. This paper provides a systematic review of this integration approach, outlining the technical evolution and current research status of autoencoder-based, clustering-based, and hybrid methods. It establishes a classification framework based on feature extraction, clustering strategies, and model fusion dimensions, while comparing the advantages, limitations, and applicable scenarios of various approaches. Key bottlenecks identified include the misalignment between feature learning objectives and clustering goals, as well as low model coupling efficiency. Based on domain-specific requirements, the study proposes future research directions emphasizing lightweight architecture, fine-grained processing, and enhanced robustness. This review offers a clear theoretical framework and methodological guidance for related research, providing a basis for selecting appropriate implementation approaches.