A few-shot learning method for single-object visual anomaly detection

We propose a few-shot learning method for visually inspecting single objects in an industrial setting. The proposed method is able to identify whether or not an object is defective by comparing its visual appearance with a small set of images of the “working” object, i.e., the object that passes the visual inspection. The method does not require images of defective objects. Furthermore, the method does not need to be “trained” when used to inspect new, previously unseen, objects. This suggests that the method can be easily deployed in industrial settings. We have evaluated the method on three visual anomaly detection benchmarks—1) MVTec, 2) MPDD, and 3) VisA. On the first two datasets the proposed method achieves performance that is comparable to state-ofthe- art methods that require access to object-specific training data. Model performance on VisA is poor; however, it is to be noted that the model was never trained on VisA dataset. We also show that the proposed model boasts fast inference times, which is a plus for industry applications. This project is funded in part by Axiom Plastics Inc., and we have evaluated the proposed method on a proprietary dataset provided by Axiom. The results confirm that the proposed method is well-suited for single-object visual anomaly detection in industry settings.

Name thesis:degree_name

Master of Science (MSc)

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Ontario Institute of Technology

Year dc:date.issued

2024

Author dc:creator

• Ejaz, Neha

Advisor dc:contributor.advisor

• Qureshi, Faisal

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10155/1777

OAI identifier oai:identifier

oai:ontariotechu.scholaris.ca:10155/1777