Polunadzirano učenje segmentacijskih modela zasnovanih na pažnji

master's thesis

• Title (english): Semisupervised learning of segmentation models based on attention
• Author: Matej Škrabić
• Mentor: Marko Čupić (mentor)
• Committee member: Marko Čupić (predsjednik povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Committee member: Željka Mihajlović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-09-14, Croatia

Abstract

U okviru rada predstavljeni su različiti pristupi za segmentaciju slike, uključujući tradicionalne algoritme, konvolucijske modele za semantičku segmentaciju i modele zasnovane na pažnji. Detaljno je opisana arhitektura odabranih modela: Transformer, Vision Transformer, Data-efficient Image Transformer i Segmenter. Kako bi se ublažila potreba za velikom količinom slika označenih na razini piksela, istražene su različite metode polunadziranog učenja s naglaskom na metode konzistencijskog učenja. Implementirana je metoda polunadziranog učenja temeljena na modelu učenik-učitelj uz fotometrijske perturbacije i upotrebu maskirajućeg autoenkodera te je iskorištena za učenje modela Segmenter na javno dostupnom skupu podataka CityScapes.

Abstract (english)

This thesis describes different image segmentation techniques, including traditional ad-hoc algorithms and machine learning approaches: convolutional and attention-based models. We described and explained the architecture of the following models: Transformer, Vision Transformer, Data-efficient Image Transformer and Segmenter. Semi-supervised training is utilized to tackle the problematic need for a large amount of pixel-wise labeled data in the context of semantic segmentation. Focus is placed on consistency regularization methods. We implemented a semi-supervised training procedure based on the student-teacher model with the use of photometric perturbations and token masking, and used it for training of the Segmenter model on the publicly available CityScapes dataset.

• Keywords: strojno učenje; računalni vid; semantička segmentacija; polunadzirano učenje; modeli zasnovani na pažnji
• Keywords (english): machine learning; computer vision; semantic segmentation; semi-supervised learning; attention-based models
• Language: croatian
• URN:NBN: urn:nbn:hr:168:837960
• 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: 2024-03-28 07:55:13