A Review of Industrial Image Anomaly Detection Based on Unsupervised Deep Learning

Abstract

Under industry 4.0, automated quality inspection, aiming to automatically identify surface defects on product, is vital for intelligent manufacturing. In the real-world industrial scenarios, the lack of defective samples and the high cost of manual annotation pose significant obstacles to the effective application of supervised learning methods. Unsupervised deep learning uses only normal samples to train the model, which provides a viable approach. This review surveys unsupervised deep learningbased methods for industrial image anomaly detection. First, the necessity of unsupervised algorithms and the main challenges they face were demonstrated. Existing approaches are classified into five categories: reconstruction-based, teacher–student network-based, feature memory repositorybased, generative network-based, and vision–language model-based. For each category, I discuss its underlying mechanism, advantages, failure scenarios, and applicable tasks, and further provide a comparative analysis among them. Then, common benchmark datasets such as MVTec AD and evaluation metrics such as AUROC and PRO were introduced, and the influence of datasets and metrics on method evaluation was discussed. Finally, open challenges such as cross-domain generalization and real-time constraints were summarized, and future research trends were prospected from directions such as multimodal fusion and lightweight deployment.