Details

Anomaly detection in uncontrolled industrial environments with data preprocessing
ID Boljević, Luka (Author), ID Skočaj, Danijel (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (9,84 MB)
MD5: E4F4B0D4728F85215EEBC4AC5F2ECD35

Abstract
Anomaly detection in industrial applications is commonly evaluated on standard benchmark datasets such as MVTec AD and VisA, whose data were captured under controlled conditions. Performance on these benchmarks does not necessarily transfer to real-world environments. In uncontrolled industrial settings, images often contain multiple objects, varying backgrounds, and inconsistent orientations, which significantly complicates anomaly detection. This thesis addresses these challenges by proposing a three-stage anomaly detection pipeline designed to reduce environmental effects prior to anomaly detection. The pipeline consists of object detection, canonicalization, and anomaly detection, transforming each object into a canonical representation. The approach is evaluated on an electrical insulator defect detection dataset, as well as select categories from the MVTec AD and VisA benchmarks, using unsupervised anomaly detection methods. Experimental results demonstrate that appropriate preprocessing consistently improves anomaly detection performance on the insulator dataset, with measurable improvements on selected MVTec AD and VisA categories.

Language:English
Keywords:Computer vision, anomaly detection, data preprocessing, industrial environment, uncontrolled environment
Work type:Master's thesis/paper
Organization:FRI - Faculty of Computer and Information Science
Year:2026
PID:20.500.12556/RUL-182384 This link opens in a new window
Publication date in RUL:08.05.2026
Views:29
Downloads:6
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Detekcija anomalij v nenadzorovanih industrijskih okoljih s predobdelavo podatkov
Abstract:
Detekcija anomalij v industrijskih aplikacijah se običajno vrednoti na standardnih podatkovnih množicah, kot sta MVTec AD in VisA, katerih podatki so bili zajeti v nadzorovanih pogojih. Uspešnost metod na teh podatkovnih množicah se ne prenese nujno v realna okolja. V nenadzorovanih industrijskih okoljih slike pogosto vsebujejo več objektov, raznolika ozadja in nekonsistentne orientacije, kar bistveno otežuje detekcijo anomalij. To magistrsko delo rešuje te izzive s tristopenjskim postopkom za detekcijo anomalij, katerega cilj je zmanjšanje vpliva okolja pred samo detekcijo anomalij. Postopek vključuje detekcijo objektov, kanonizacijo in detekcijo anomalij, pri čemer se vsak objekt pretvori v kanonično obliko. Predlagan pristop je ovrednoten na podatkovni množici za detekcijo anomalij na električnih izolatorjih ter na izbranih kategorijah referenčnih podatkovnih množic MVTec AD in VisA, pri čemer so uporabljene nenadzorovane metode detekcije anomalij. Eksperimentalni rezultati kažejo, da ustrezna predobdelava dosledno izboljša uspešnost detekcije anomalij na podatkovni množici z izolatorji ter ima merljiv vpliv tudi na izbrane kategorije podatkovnih množic MVTec AD in VisA.

Keywords:Računalniški vid, detekcija anomalij, predobdelava podatkov, industrijsko okolje, nenadzorovano okolje

Similar documents

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

Back