prikaz prve stranice dokumenta Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije
No public access
undergraduate thesis
Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije
Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing, 2024. urn:nbn:hr:168:657616
Runtas, Jurica
University of Zagreb
Faculty of Electrical Engineering and Computing

Institutional repository: FER Repository

Cite this document

Runtas, J. (2024). Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije (Undergraduate thesis). Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing. Retrieved from https://urn.nsk.hr/urn:nbn:hr:168:657616

Runtas, Jurica. "Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije." Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2024. https://urn.nsk.hr/urn:nbn:hr:168:657616

Runtas, Jurica. "Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije." Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2024. https://urn.nsk.hr/urn:nbn:hr:168:657616

Runtas, J. (2024). 'Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije', Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, accessed 14 January 2025, https://urn.nsk.hr/urn:nbn:hr:168:657616

Runtas J. Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije [Undergraduate thesis]. Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing; 2024 [cited 2025 January 14] Available at: https://urn.nsk.hr/urn:nbn:hr:168:657616

J. Runtas, "Nesigurnost segmentacije slike s obzirom na uzorke izvan distribucije", Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, 2024. Available at: https://urn.nsk.hr/urn:nbn:hr:168:657616

Metadata
TitleNesigurnost segmentacije slike s obzirom na uzorke izvan distribucije
Title (english)Uncertainty of Image Segmentation Regarding Out-of-distribution Samples
AuthorJurica Runtas
MentorTomislav Petković (mentor)
Committee memberTomislav Petković (predsjednik povjerenstva)
Committee memberMarko Subašić (član povjerenstva)
Committee memberDonik Vršnak (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb
Defense date and country2024-07-03, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Abstract
Semantička segmentacija je zadatak računalnog vida u kojem se svaki piksel slike dodjeljuje u jednu od unaprijed definiranih klasa. Duboke neuronske mreže (DNM) su suvremeno rješenje tog zadatka i obično su trenirane tako da rade na zatvorenom skupu unaprijed definiranih klasa. U okruženjima otvorenog skupa, moguće je susresti anomalije, tj., semantički nepoznate objekte. Cestovna vožnja je primjer takvog okruženja u kojem je, sa stajališta sigurnosti, bitno osigurati da DNM može indicirati da operira izvan svoje naučene semantičke domene. Jedna metoda koja se koristi za tu svrhu je maksimizacija entropije uparena s korakom postprocesiranja zvanim meta klasifikacija koji povećava pouzdanost detekcije anomalija. U ovom radu, predlažemo pristup meta klasifikaciji koji značajno povećava pouzdanost detekcije anomalija u odnosu na originalni pristup tako što zamjenjuje logističku regresiju s jednostavnom neuronskom mrežom kao meta klasifikatorom. Analiziramo njegovu performansu i dajemo pregled potencijalnih arhitektura, prednosti i nedostataka korištenja neuronske mreže umjesto logističke regresije. Nadalje, dodatno evaluiramo metodu maksimizacije entropije i ukazujemo na to da je potreban oprez prilikom korištenja te metode u praksi kako bi se osigurala njena učinkovitost.
Abstract (english)
Semantic segmentation is a computer vision task in which each pixel of an image is assigned into one of the predefined classes. Deep neural networks (DNNs) are a contemporary solution for such task and are usually trained to operate on a predefined closed set of classes. In open-set environments, it is possible to encounter anomalies, i.e., semantically unknown objects. Road driving is an example of such environment in which, from a safety standpoint, it is important to ensure that a DNN is able to indicate when it is operating outside of its learned semantic domain. One of the methods used for that purpose is entropy maximization paired with a post-processing step called meta classification which increases the reliability of anomaly detection. We propose a meta classification approach that significantly improves the reliability of anomaly detection in comparison to the original approach by substituting logistic regression meta classifier with a lightweight neural network meta classifier. We analyze its performance and examine potential model architectures, advantages and drawbacks of using it instead of logistic regression. Furthermore, we additionally evaluate the entropy maximization and show that caution must be taken when using it in practice in order to ensure its effectiveness.
Keywords
Keywords (english)
Languageenglish
URN:NBNurn:nbn:hr:168:657616
Study programmeTitle: Computing
Study programme type: university
Study level: undergraduate
Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsClosed access
Terms of useLicence In copyright icon
Public note
Created on2024-10-12 22:02:54