notesum.ai

Published at November 25

FUN-AD: Fully Unsupervised Learning for Anomaly Detection with Noisy Training Data

cs.CV

Released Date: November 25, 2024

Authors: Jiin Im¹, Yongho Son², Je Hyeong Hong¹

Aff.: ¹Dept. Electronic Engineering, Hanyang University; ²Dept. Artificial Intelligence, Hanyang University

Arxiv: http://arxiv.org/abs/2411.16110v1

Refer to caption

Type			One-class classification					Fully unsupervised
Dataset	Method	CS-Flow [29]	PaDiM [9]	PatchCore [28]	SimpleNet [22]	RealNet [49]	SoftPatch [38]	InReaCh [25]	FUN-AD (Ours)
MVTec AD	No overlap	93.8 / -	94.5 / 96.6	98.1 / 94.7	97.7 / 95.3	99.2 / 96.9	98.3 / 97.3	92.2 / 96.9	99.2 / 98.6
MVTec AD	Overlap	65.5 / -	70.1 / 91.4	81.6 / 70.8	59.3 / 52.9	96.8 / 96.0	98.3 / 95.1	92.4 / 97.2	98.8 / 98.6
VisA	No overlap	82.3 / -	85.8 / 98.3	89.8 / 95.7	90.4 / 96.7	93.5 / 98.7	89.6 / 98.5	83.8 / 97.6	95.0 / 99.2
VisA	Overlap	64.3 / -	70.1 / 92.3	82.3 / 81.6	50.4 / 57.9	92.4 / 97.7	89.5 / 96.3	78.5 / 93.6	95.1 / 99.2