Kombiniranje modelov razvrščanja vzorcev za razpoznavanje čustvenih izrazov iz posnetkov obraza v nekontroliranem okolju

diplomsko delo

Juš Osojnik (Author), Božidar Potočnik (Mentor), Martin Šavc (Co-mentor)

Abstract

V tem delu smo se ukvarjali z razpoznavanjem čustvenih izrazov v nekontroliranem okolju. Uporabljali smo metodo prenosnega učenja, kjer smo učili arhitekture konvolucijskih nevronskih mrež: EfficientNetB0, ResNet50, DenseNet121, InceptionV3 in Xception, na naboru podatkovnih zbirk FER-2013, AffectNet, AFEW/SFEW in Aff-Wild2. Modele smo nato kombinirali na osnovi rezultatov z metodama povprečenja in glasovanja. Modele smo kombinirali tudi na osnovi izluščenih značilnic. Uspešnost modelov smo merili po metrikah natančnosti in ocene F1. Na podatkovni zbirki FER-2013 smo dosegli najboljšo natančnost 72 %, na zbirkah AffectNet 67 %, AFEW/SFEW 47 % in Aff-Wild2 52 % natančnost. Z našimi rezultati smo se približali najuspešnejšim raziskavam, ki so na posameznih podatkovnih zbirkah dosegle natančnosti: FER-2013 77 %, AffectNet 67 %, AFEW/SFEW 54 % in Aff-Wild2 52 %.

Keywords

prepoznavanje čustvenih izrazov;slike obrazov;globoke nevronske mreže;modelno združevanje;ekstrakcija značilnic;okolje Keras;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2023
Typology:	2.11 - Undergraduate Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[J. Osojnik]
UDC:	004.93(043.2)
COBISS:	174792451
Views:	138
Downloads:	16
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Combining classification models for facial emotion recognition in uncontrolled environments
Secondary abstract:	In this work, we focused on facial emotion recognition in an uncontrolled environment. We used the transfer learning method, where we trained the architectures of convolutional neural networks: EfficientNetB0, ResNet50, DenseNet121, InceptionV3, Xception, on the datasets FER-2013, AffectNet, AFEW/SFEW, and Aff-Wild2. We then combined the models based on the results using both averaging and voting methods. We also combined the models based on extracted features. The performance of the models was measured using accuracy and F1 score metrics. On the FER-2013 dataset, we achieved an accuracy of 72 %, on AffectNet 67 %, on AFEW/SFEW 47 %, and on Aff-Wild2 52 %. In our research we aproached the state of the art results achieved on the individual datasets, which are: FER-2013 77 %, AffectNet 67 %, AFEW/SFEW 54 % and Aff-Wild2 52 %.
Secondary keywords:	facial emotion recognition;facial images;deep neural networks;model;ensemble;feature extraction;keras applications;
Type (COBISS):	Bachelor thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Računalništvo in informacijske tehnologije
Pages:	1 spletni vir (1 datoteka PDF (XV, 78 f.))
ID:	19924208