Institutional Repository of Vilnius University
Nuostolio funkcijų taikymo vaizdų panašume tyrimas /
Abstract
dc:descriptionTogether with new technologies came the increase in the amounts of readily available visual data, creating the need to be able to find, recognize, and compare images using not only textual but also visual queries. A widely researched method for solving such image similarity comparison problems is artificial neural networks. This study focuses on loss functions – an integral part in the training process of such networks. The aim of this work is to research the loss functions used in the field of image similarity, determine which functiontrained models best solve image recognition and retrieval tasks, and analyse how model performance depends on the use of data mining. Models based on different depths of ResNet architecture neural networks were trained using a subset of the “Google Landmarks Dataset v2” (GLDv2) dataset. For model evaluation the “Oxford Buildings” and “Paris Dataset” together with “Revisited Oxford” and “Revisited Paris” datasets, were used. The best mean average precision (mAP) scores, without using online pair mining, were achieved by networks trained with the CosFace loss function. The use of online pair mining improved the models’ performance. The overall best results were obtained by models trained using online pair mining and the SphereFace loss function.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gakaitė, Gabija,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD210645545&prefLang=en_US