Unsupervised Defect Clustering from Optical One-Class Anomaly Detection

This paper introduces a fully automated pipeline for defect clustering aimed at reducing manual labeling effort for AI-based classification. The pipeline uses anomaly regions of objects, acquired through one-class anomaly detection, to generate feature vectors based on DINOv2 patch feature vectors. The high performance of DINOv2's patch feature vectors in describing local features eliminates the need for additional AI model training in the pipeline. The pipeline was evaluated using defective images from the MVTec AD dataset.