No public access
undergraduate thesis
10   0
Deep neural architectures for object localization
Matija Folnović (2015)
University of Zagreb
Faculty of Electrical Engineering and Computing
Metadata
TitleDuboke neuronske arhitekture za lokalizaciju objekata
AuthorMatija Folnović
Mentor(s)Siniša Šegvić (thesis advisor)
Abstract
Konvolucijske neuronske mreže postižu vrhunske rezultate na području klasifikacije slika. Međutim, te uspjehe nije jednostavno preslikati na područje lokalizacije objekata u slikama. U ovom radu razmatramo algoritme za lokalizaciju objekata u slikama koristeći konvolucijske neuronske mreže. Osim toga, razmatramo translacijski autoenkoder kao rješenje problema koji konvolucijska neuronska mreža ima s translacijom. U konačnici, razmatramo algoritam za lokalizaciju koristeći translacijski autoenkoder. Prikazani su i komentirani dobiveni rezultati klasifikacije slika u skupu MNIST, kao i lokalizacije slika iz istog skupa.
Keywordsneural networks machine learning computer vision object localization object classification convolutional neural networks translational autoencoder capsules
Parallel title (English)Deep neural architectures for object localization
Committee MembersSiniša Šegvić (committee chairperson)
Siniša Šegvić (committee chairperson)
Mladen Karan (committee member)
Marin Golub (committee member)
Goran Glavaš (committee member)
Leonardo Jelenković (committee member)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
PlaceZagreb
StateCroatia
Scientific field, discipline, subdisciplineTECHNICAL SCIENCES
Computing
Study programme typeuniversity
Study levelundergraduate
Study programmeElectrical Engineering and Information Technology
Academic title abbreviationuniv. bacc. ing. el. techn. inf.
Genreundergraduate thesis
Language Croatian
Defense date2015-07-13
Parallel abstract (English)
Convolutional neural networks achive excellent results in the field of image classification. However, this success is not easily mapped to the field of object localization. In this paper, we consider algorithms for object localizations using convolutional neural networks. Furthermore, we consider translational autoencoder as a solution for problems convolutional neural network has with translations. Finally, we consider the algorithm for object localization using translational autoencoder. Results of image classification and object localizations of images in MNIST dataset are shown and discussed.
Parallel keywords (Croatian)neuronske mreže strojno učenje računalni vid lokalizacija objekata klasifikacija objekata konvolucijske neuronske mreže translacijski autoenkoder kapsule
Resource typetext
Access conditionNo public access
Terms of usehttp://rightsstatements.org/vocab/InC/1.0/
URN:NBNhttps://urn.nsk.hr/urn:nbn:hr:168:771062