diplomsko delo
Klemen Štefe (Author), Žiga Emeršič (Mentor)

Abstract

Najbolj ključen dejavnik pri delovanju globokih biometričnih modelov je njihov postopek učenja. Diplomsko delo raziskuje različne pristope k optimizaciji učenja globokih nevronskih mrež, z namenom izboljšave njihove klasifikacijske točnosti. Osredotočamo se na metode iz področja zmanjšanja prileganja podatkom in vpliva različnih hiperparametrov na rezultate učenja. Za raziskavo uporabimo modele naučene na podatkovni zbirki ImageNet, ki jih s pomočjo prenosnega učenja prilagodimo za klasifikacijo ljudi na podlagi njihovega uhlja. Zaradi vpliva strojne opreme, testiramo tudi čas učenja posameznih modelov, ter povprečne hitrosti njihovih napovedi. Ugotavljamo, da je za našo učno množico najbolj primeren model ResNet18, z najvišjo točnostjo 56 odstotkov, sledi pa mu GoogLeNet z 51 odstotki.

Keywords

prenosno učenje;prileganje podatkov;regularizacija;augmentacija;ImageNet;razpoznava uhljev;biometrija uhljev;visokošolski strokovni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [K. Štefe]
UDC: 004.93:57.087.1(043.2)
COBISS: 162224131 Link will open in a new window
Views: 53
Downloads: 20
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Optimizing the training of deep biometric models in ear biometrics
Secondary abstract: The key factor in the performance of deep biometric models lies in their learning process. This study investigates different approaches to optimize the learning of deep neural networks, aiming to enhance their classification accuracy. We focus on methods that reduce overfitting and examine the impact of various hyperparameters. To conduct this research, we utilize models trained on the ImageNet dataset, which we fine-tune using transfer learning to classify people based on their ears. Furthermore, we assess the training times and average prediction speeds of individual models, considering hardware constraints. The results show that ResNet18 is the most suitable model for our training data, achieving best accuracy of 56%, closely followed by GoogLeNet with 51%
Secondary keywords: neural networks;classification;transfer learning;data fitting;regularization;augmentation;ImageNet;ear recognition;ear biometrics;computer science;diploma;Biometrija;Biometrična identifikacija;Nevronske mreže (računalništvo);Računalniški vid;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000470
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 46 str.
ID: 21439461