Konvolucijski modeli za semantičku segmentaciju primjeraka

master's thesis

• Title (english): Convolutional models for semantic segmentation of object instances
• Author: Bruno Kovač
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Željka Mihajlović (član povjerenstva)
• Committee member: Tomislav Hrkać (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Lokalizacija i segmentacija primjeraka objekata jedni su od bitnih problema dubokog
učenja. Mask R-CNN predstavlja stanje tehnike modela koji rješavaju te zadatke.
U sklopu rada pojašnjene su sve komponente Mask R-CNN-a poput kralježnice u vidu
rezidualnih neuronskih mreža ResNet, rezolucijske piramide značajki FPN, modula za
predlaganje regija interesa RPN te samih glava za klasifikaciju, okvire i semantičku
segmentaciju specifičnih za cijeli model. Model Mask R-CNN-a implementiran je i
naučen na dva različita skupa podataka PASCAL VOC-2012 i AOLP. Prikazane su dobiveni
vrijednosti AP metrika te sami rezultati na slikama gdje su objekti klasificirani,
lokalizirani i segmentirani. Za oba skupa podataka i područja primjene također su prikazani
i rezultati na slikama izvan navedenih skupova kako bismo prikazali mogućnost
generalizacije našeg modela.

Abstract (english)

Object detection and instance segmentation are one of the most important problems
in deep learning. Mask R-CNN represents state-of-the-art model which solves these
tasks. As part of the paper, all components of Mask R-CNN are explained, such as
backbone in form of residual neural networks ResNet, feature pyramid network FPN,
region proposal network RPN and classification, bounding box and semantic segmentation
heads specific for the whole model. Mask R-CNN model was implemented and
trained on two different datasets PASCAL VOC-2012 and AOLP. Values of AP metrics
are shown and also results on images where objects are classified, localized and
segmented. For both datasets and areas of application results are also shown on images
not included in mentioned datasets to show generalization of our model.

• Keywords: Konvolucijske neuronske mreže; ResNet; rezolucijska piramida značajki; FPN; modul za predlaganje regija interesa; RPN; Mask R-CNN; klasifikacija; lokalizacija; detekcija; segmentacija primjeraka; VOC-2012; AOLP
• Keywords (english): Convolutional neural networks; ResNet; Feature Pyramid Network; FPN; Region Proposal Network; RPN; Mask R-CNN; classification; localization; detection; instance segmentation; VOC-2012; AOLP
• Language: croatian
• URN:NBN: urn:nbn:hr:168:984466
• 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
• Repository: FER Repository
• Created on: 2023-01-13 12:44:26