Polunadzirana semantička segmentacija utemeljena na pseudooznačavanju

master's thesis

• Title (english): Semi-supervised semantic segmentation based on pseudolabelling
• Author: Toma Petrač
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Vladimir Čeperić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Semantička segmentacija prirodnih scena važan je zadatak računalnog vida s mnogim zanimljivim primjenama. Jako dobre rezultate na tom zadatku postižu konvolucijski modeli s lakom klasifikacijskom osnovom i ljestvičastim naduzorkovanjem. Ovaj rad bavi se polunadziranim učenjem takvih modela, što znači da model uči na grupama koje se sastoje od označenih i neoznačenih slika. Koristi se polunadzirana metoda pseudooznačavanja neoznačenih slika, najprije na manjem i jednostavnijem skupu podataka, a zatim na nešto kompleksnijem. Provedeno je mnogo iscrpnih eksperimenata koji su pokazali da metoda doista daje bolje rezultate nego uobičajeni postupak nadziranog učenja.

Abstract (english)

The segmentation of natural scenes is an important task of computer vision with many interesting applications. Very good results on this task are achieved by convolutional models with a light-weight general purpose architecture as the main recognition engine, and upsampling with lateral connections. This paper deals with the semi-supervised learning of such models, which means that the model learns in groups consisting of both labeled and unlabeled images. A semi-supervised method of pseudo-labeling is used, first on a smaller and simpler data set, and then on a somewhat more complex one. Many exhaustive experiments have been conducted which have shown that the method really gives better results than the usual supervised learning procedure.

• Keywords: računalni vid; konvolucijski modeli; ljestvičasto naduzorkovanje; polunadzirano učenje; semantička segmentacija; pseudooznake; CIFAR-10; CamVid
• Keywords (english): computer vision; convolutional neural networks; ladder-style upsampling; semi-supervised learning; pseudolabelling; semantic segmentation; CIFAR-10; CamVid
• Language: croatian
• URN:NBN: urn:nbn:hr:168:106671
• 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: 2023-01-13 13:28:28