Mixed supervision for surface-defect detection

Božič, Jakob (author), Skočaj, Danijel (mentor), Tabernik, Domen (co-mentor)

Surface-defect detection aims to identify defective regions in images. Recently, various deep-learning based solutions have been proposed to tackle this task, requiring different types of labels for training data. Fully supervised approaches require costly-to-produce pixel-level labels for all samples, but deliver excellent performance. On the other end of the spectrum, unsupervised methods are trained from normal data only, but often fall short in performance. Neither of these are capable of utilizing all the available data, as the first can not employ weakly labeled data and the latter fail to consider defective samples. We introduce mixed supervision to bridge the performance gap between fully supervised and unsupervised methods by enabling learning from all available data. A fully supervised method that can learn from weakly-labeled data is proposed and an unsupervised method is extended to extract knowledge from defective samples. Extensive evaluation of mixed supervision shows that allowing both methods to learn from previously unused data significantly improves their performance. In a detailed analysis of robustness, unsupervised methods prove to be surprisingly robust to false negatives in the training data, showing potential for use in fully unsupervised scenario with completely unlabeled data.

• Language: English
• Keywords: deep learning, surface-defect detection, mixed supervision, anomaly detection, robustness
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-141253
• COBISS.SI-ID: 124810243
• Publication date in RUL: 27.09.2022

Secondary language

• Language: Slovenian
• Title: Učenje z mešanim nadzorom za detekcijo površinskih napak

Abstract

Detekcija površinskih napak se ukvarja z iskanjem nepravilnosti v slikah. V zadnjih letih je bilo za reševanje tega problema predlaganih veliko rešitev, ki temeljijo na globokem učenju in ki za učenje zahtevajo podatke z različno natančnostjo oznak. Polno nadzorovane metode dosegajo odlične rezultate vendar za učenje potrebujejo za vse slike oznake na nivoju sl. elem., ki so zahtevne za pridobitev. Nasprotno, nenadzorovane pristope učimo le na normalnih slikah, a pogosto ne delujejo tako dobro. Nobeden izmed teh pristopov za učenje ne more uporabiti vseh razpoložljivih podatkov, saj prvih ne moremo učiti s podatki, označenimi na nivoju slik, druge pa ne izkoriščajo slik z napakami. V tem delu vpeljemo učenje z mešanim nadozorom, ki omogoča uporabo vseh podatkov in tako zmanjšuje razlike med pristopoma. Predlagamo polno nadzorovano metodo, ki jo lahko učimo z označbami na nivoju slik in razširimo nenadzorovano metodo z uporabo slik z napakami. Obsežno vrednotenje učenja z mešanim nadzorom pokaže, da se delovanje obeh vrst metod izboljša z uporabo vseh razpoložljivih podatkov. Analiza robustnosti pokaže, da so nenadzorovane metode presenetljivo neobčutljive na prisotnost slik z napakami v učni množici in tako potencialno vodijo k učenju v polno nenadzorovanem načinu, kjer podatki nimajo nobenih oznak.

• Keywords: globoko učenje, detekcija površinskih napak, mešani nadzor, detekcija anomalij, robustnost