How Similar is Too Different? A DINOv2-Based Defect Detection Method
Fulltext URI
Document type
Additional Information
Date
relationships.isAuthorOf
Journal Title
Journal ISSN
Volume Title
Source
Publisher
Abstract
Problem-free defect detection with neural networks is still a challenge in flexible industrial environments, especially in contract manufacturing. A training-free defect detection system using DINOv2 is proposed to address these challenges of resource-intensive model training in flexible industrial environments. The method leverages patch-level embeddings, where each patch represents a small image region, and uses cosine similarity to detect and localize anomalies without prior training. Two strategies are compared: Method I (all patches) and Method II (least similar patches). Experiments on MVTec AD, VisA, and KSDD2 show that Method II outperforms Method I, with Area Under the Curve improvements of 6 − 8% and achieved results of 91.3%, 82.8%, and 92.8%, respectively. The study underscores the significance of reference image diversity, object orientation, and patch selection. The approach’s simplicity and adaptability make it a scalable solution for industrial visual inspection, eliminating the need for extensive labeled data and individualized model training.
