SuperSimpleNet

detekcija anomalij z uporabo preproste globoke arhitekture in učenja z mešanim nadzorom

Blaž Rolih (Author), Danijel Skočaj (Mentor), Matic Fučka (Co-mentor)

Abstract

Cilj detekcije anomalij je prepoznati in lokalizirati anomalne regije na zajetih objektih, kar je v sodobnih proizvodnih procesih ključni element zagotavljanja kakovosti. Kljub temu številni obstoječi pristopi pogosto ne izpolnjujejo vseh zahtev, ki jih postavlja industrija. Te zajemajo zahtevo po konsistentnosti, hitrem delovanju in visoki uspešnosti, vključno z zmožnostjo učinkovitega izkoriščanja razpoložljivih učnih podatkov. V tem delu predstavimo novo diskriminativno metodo SuperSimpleNet, ki smo jo razvili iz metode SimpleNet, s katero naslovimo navedene pomanjkljivosti. Metoda SuperSimpleNet izboljšuje uspešnost detekcije, hitrost napovedi in stabilnost učenja ter je ena izmed zelo redkih metod, ki hkrati omogoča delovanje v nenadzorovanem, šibkem, mešanem in polno nadzorovanem scenariju. To omogoča uporabo vseh razpoložljivih podatkov. Poleg tega uvedba aktivnega učenja pripomore k pametni izbiri novih primerov za označevanje, kar dodatno optimizira proces. Metoda SuperSimpleNet z mero AUC 98,0 % na podatkovni množici SensumSODF presega vse dosedanje metode in hkrati dosega vrhunske rezultate z AP 97,8 % na KolektorSDD2, AUC 93,6 % na VisA in AUC 98,3 % na MVTec AD. Je tudi ena izmed najhitrejših metod s časom napovedi 9,5 ms in prepustnostjo 262 slik na sekundo.

Keywords

detekcija anomalij;odkrivanje napak;mešani nadzor;računalništvo in informatika;magisteriji;

Data

Language:	Slovenian
Year of publishing:	2024
Typology:	2.09 - Master's Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[B. Rolih]
UDC:	004.93:004.85(043.2)
COBISS:	210256387
Views:	117
Downloads:	60
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	SuperSimpleNet: anomaly detection using a simple deep architecture and learning with mixed supervision
Secondary abstract:	The goal of anomaly detection is to identify and localize anomalous regions in captured objects, a key element of ensuring quality in modern manufacturing processes. However, many existing approaches often fail to meet all the industry's requirements, which include consistency, fast operation, and high performance, as well as the ability to effectively utilize all the available training data. In this work, we present a novel discriminative method, SuperSimpleNet, which evolved from SimpleNet, that addresses these shortcomings. SuperSimpleNet enhances detection performance, inference speed, and training stability, while it's one of the very few methods that supports operation in unsupervised, weak, mixed, and fully supervised setting. This allows for the utilization of all available data. Moreover, the introduction of active learning aids with the smart selection of new samples for annotation, further optimizing the process. SuperSimpleNet surpasses all existing methods with AUC of 98.0 % on SensumSODF dataset and achieves competitive results with AP 97.8 % on KolektorSDD2, AUC 93.6 % on VisA, and AUC 98.3 % on MVTec AD. It is also one of the fastest methods, with inference time of 9.5 ms and throughput of 262 images per second.
Secondary keywords:	computer vision;anomaly detection;deep learning;mixed supervision;computer science;computer and information science;master's degree;Računalniški vid;Globoko učenje (strojno učenje);Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Master's thesis/paper
Study programme:	1000471
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	1 spletni vir (1 datoteka PDF (92 str.))
ID:	24920736