Details

Unsupervised reconstructive and discriminative methods for surface anomaly detection and localisation
ID Zavrtanik, Vitjan (Author), ID Skočaj, Danijel (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (17,87 MB)
MD5: 738F1E292D4796CF21FC0D9B1ABC4444

Abstract
This thesis deals with the task of unsupervised surface anomaly detection and localization. Through an overview of the surface anomaly detection field, it describes the main paradigms of the current surface anomaly detection methods, analyses the field's most recent best performing methods and points out their shortcomings. As the main part of the thesis, several methods addressing the issues of the reconstructive and the discriminative anomaly detection paradigms are proposed. The three proposed surface anomaly detection methods represent the main contributions to science of the doctoral work. The RIAD method belongs to the reconstructive paradigm. It reformulates image reconstruction as an iterative inpainting process, preventing the reconstruction of anomalies, making them detectable. The DRAEM method is a discriminative anomaly detection method that is trained using simulated anomalies. It proposes a reconstructive and discriminative framework that removes anomalies from the image and then accurately localizes them based on the difference between the input image and the anomaly-free reconstruction. The DSR method simulates anomalies directly in the discrete latent space, which improves the diversity of simulated anomalies and enables the accurate detection of real anomalies. As an additional contribution, the use of the proposed discriminative methods on other data modalities has been explored, demonstrating the significant adaptability of the proposed architectures.

Language:English
Keywords:anomaly detection, unsupervised learning, surface anomaly detection
Work type:Doctoral dissertation
Typology:2.08 - Doctoral Dissertation
Organization:FRI - Faculty of Computer and Information Science
Year:2025
PID:20.500.12556/RUL-175005 This link opens in a new window
COBISS.SI-ID:254035715 This link opens in a new window
Publication date in RUL:13.10.2025
Views:192
Downloads:50
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Nenadzorovane rekonstrukcijske in diskriminativne metode za detekcijo in lokalizacijo površinskih anomalij
Abstract:
Ta disertacija obravnava nenadzorovano detekcijo in lokalizacijo anomalij na površinah. V disertaciji so predstavljene glavne paradigme trenutnih detekcijskih metod, analizirane trenutne najboljše metode, izpostavljene pa so tudi njihove pomanjkljivosti. Delo predlaga metode, ki naslavljajo probleme rekonstrukcijske in diskriminativne paradigme detekcije anomalij. V disertaciji so, kot glavni prispevek k znanosti, predstavljene tri metode za detekcijo površinskih anomalij na slikah. Metoda RIAD pripada rekonstrukcijski paradigmi in predlaga izvajanje rekonstrukcije s postopkom iterativnega vrisovanja. Metoda DRAEM je diskriminativna metoda, ki je učena le z uporabo simuliranih anomalij. Sestavljena je iz rekonstrukcijskega dela in diskriminativnega dela. Rekonstrukcijski del anomalije implicitno zazna in jih iz slike izbriše. Diskriminativni del anomalije natančno lokalizira na podlagi naučene implicitne funkcije razdalje med vhodno sliko in njeno rekonstrukcijo. Diskriminativna metoda DSR za učenje uporablja anomalije simulirane v naučenem diskretnem prostoru. To izboljša raznolikost simuliranih anomalij in omogoča natančno detekcijo pravih anomalij. Dodatni prispevki znanosti te disertacije so adaptacije predlaganih diskriminativnih metod na druge podatkovne modalnosti, kot sta zvok in globinske slike. Doseženi rezultati kažejo na izjemno prilagodljivost predlaganih arhitektur.

Keywords:detekcija anomalij, nenadzorovano učenje, detekcija anomalij na površinah

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back