Anomaly detection with diffusion models

Fučka, Matic (author), Skočaj, Danijel (mentor), Zavrtanik, Vitjan (comentor)

Surface defect detection is a critical challenge in ensuring product quality, as defects can pose safety risks and diminish product lifespan. Unsupervised anomaly detection is a closely related problem that tries to detect anomalies without any explicit information about them during the training phase. While deep learning has introduced numerous methods, the problem has barely seen any attempts with diffusion models. To address this, we propose Transfusion, a TRANSparent difFUSION model. Our approach focuses on iterative erasure of anomalies. To achieve this we redesigned the base diffusion process and the base architecture for diffusion models. Notably, Transfusion achieves exceptional performance in anomaly detection, surpassing state-of-the-art results on VisA dataset with an impressive 98.5% AUROC and achieving competitive results on MVTec AD with an 99.2% AUROC. This advancement offers promising prospects for reliable and efficient surface defect detection in manufacturing processes.

• Language: English
• Keywords: computer vision, anomaly detection, diffusion models
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-149117
• COBISS.SI-ID: 164268803
• Publication date in RUL: 04.09.2023

Secondary language

• Language: Slovenian
• Title: Detekcija anomalij z difuzijskimi modeli

Abstract

Detekcija površinskih napak je ključni izziv pri zagotavljanju kakovosti izdelkov, saj lahko napake predstavljajo varnostno tveganje in skrajšajo življenjsko dobo izdelka. Nenadzorovana detekcija anomalij je tesno povezan problem, ki poskuša detektirati anomalije brez predhodnih informacij o njih med procesom učenja. Predlagane so bile številne metode, ki temeljijo na globokem učenju, vendar zelo malo z difuzijskimi modeli. Da bi to rešili, predlagamo model TRANSparent difFUSION ali na kratko Transfusion. Naš pristop se osredotoča na iterativno brisanje anomalij. Da bi to dosegli, smo preoblikovali osnovni difuzijski proces in osnovno arhitekturo za difuzijske modele. Transfusion dosega izjemne rezultate pri detekciji anomalij, saj z impresivno 98,5-odstotno vrednostjo mere AUROC presega vse dosedanje rezultate na podatkovni množici VisA in dosega vrhunske rezultate na podatkovni množici MVTec AD z 99,2-odstotno vrednostjo mere AUROC. Ta napredek ponuja obetavne možnosti za zanesljivo in učinkovito detekcijo površinskih napak v proizvodnih procesih.

• Keywords: računalniški vid, detekcija anomalij, difuzijski modeli