Gusta detekcija anomalija generativnim modelima

master's thesis

• Title (english): Dense anomaly detection with generative models
• Author: Anja Delić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Committee member: Leonardo Jelenković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-10, Croatia

Abstract

Standardni diskriminativni modeli za semantičku segmentaciju osjetljivi su na pojavu anomalija u ispitnim slikama. Navedeni problem rješavamo detekcijom anomalija razredom uvjetovanim normalizirajućim tokovima. Predlažemo mjere anomalnosti temeljene na teoriji informacije s probabilističkom interpretacijom te ih primjenjujemo na predikcije na razini piksela i maski.
Predložena hibridna metoda postiže kompetitivne rezultate na standardnim testnim skupovima slika i poboljšanje u odnosu na usporedive diskriminativne metode za gustu detekciju anomalija.

Abstract (english)

Standard discriminative models for semantic segmentation are sensitive to the presence of anomalies in test images. We address this issue by detecting anomalies using class-conditional normalizing flows. We define anomaly scores based on information theory with probabilistic interpretation and apply them to pixel-level and mask-level predictions. The proposed hybrid method achieves competitive results on standard benchmarks and improvement upon comparable discriminative dense anomaly detection methods.

• Keywords: semantička segmentacija; gusta detekcija anomalija; uvjetni normalizirajući tok; zajednička informacija
• Keywords (english): semantic segmentation; dense anomaly detection; class-conditional normalizing flows; mutual information
• Language: croatian
• URN:NBN: urn:nbn:hr:168:143902
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 08:42:38