Self-Supervised Learning for Industrial Visual Anomaly Detection: A Review of Recent Advances, Applications, and Open Challenges

Abstract

Industrial visual anomaly detection (VAD)—the automated identification and localization of defects, irregularities, and deviations in manufactured products—plays a critical role in ensuring product quality, operational safety, and process reliability across modern manufacturing. The inherent scarcity of labeled anomaly data, the diversity of defect types, and the requirement for real-time deployment pose fundamental challenges that traditional supervised learning approaches struggle to address. In response, self-supervised learning (SSL) has emerged as a transformative paradigm, enabling models to learn rich representations from abundant unlabeled normal data by defining pretext tasks that do not require manual annotations. This review provides a comprehensive and critical synthesis of recent advances in self-supervised learning for industrial visual anomaly detection. We examine the methodological landscape across five major SSL categories—contrastive learning, masked reconstruction, generative modeling, rotation prediction, and cross-modal pretext tasks—and map their application to key industrial domains including surface inspection, 3D component quality control, semiconductor fabrication, and predictive maintenance. A structured analysis of eight representative works—including the Iterative Mask Reconstruction Network (IMRNet), graph attention-based multivariate anomaly detection, and diffusion-enabled defect synthesis—grounds the discussion in empirical evidence. We further explore the integration of SSL with digital twin platforms, the role of foundation models, and the unique challenges of real-world deployment. Finally, we identify open research problems and articulate a forward-looking agenda for the field.