Optimizacija globokih mrež za prepoznavo čustvenih izrazov : magistrsko delo

Zupančič, Nejc

| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font

First page > Show document

Show document

Title:	Optimizacija globokih mrež za prepoznavo čustvenih izrazov : magistrsko delo
Authors:	ID Zupančič, Nejc (Author) ID Mlakar, Uroš (Mentor) More about this mentor...
Files:	MAG_Zupancic_Nejc_2025.pdf (3,15 MB) MD5: 629F6D3EBE1B3AAF3C4BD66C18AD5889
Language:	Slovenian
Work type:	Master's thesis/paper
Typology:	2.09 - Master's Thesis
Organization:	FERI - Faculty of Electrical Engineering and Computer Science
Abstract:	V sklopu magistrskega dela smo obravnavali kompleksen problem razpoznave čustvenih izrazov s pomočjo globokih nevronskih mrež. Trenutne rešitve so izredno neučinkovite, ko jih prenesemo v realni svet. V praktičnem delu smo s pomočjo evolucijskega algoritma diferencialne evolucije optimizirali parametre treh znanih arhitektur globokih nevronskih mrež – DenseNet121, ResNet50 in VGG16, z željo, da bi pridobili višjo natančnost. S pomočjo učne množice AffectNet in tehnike prenosnega učenja smo modele priredili za problem razpoznave čustvenih izrazov. Z diferencialno evolucijo smo uspešno našli hiperparametre (stopnja učenja, moment, upad uteži), ki nam dajejo malenkost boljše rezultate za posamezen model. Natančnost DenseNet121 smo zvišali za 1,72 % (z 52,24 % na 53,96 %), ResNet za 1,88 % (z 51,86 % na 53,74 %) in VGG16 za 1,59 % (z 52,01 % na 53,6 %). Prav tako smo opazovali, kako velikost množice vpliva na uspešnost. Ugotovili smo, da nam množice, velike od 30 % do 40 %, dajejo rezultate, ki so v povprečju okoli 2 % slabši od rezultatov, pri katerih smo uporabili celotno učno množico.
Keywords:	prepoznavanje čustev, konvolucijske nevronske mreže, optimizacija parametrov učenja
Place of publishing:	Maribor
Place of performance:	Maribor
Publisher:	[N. Zupančič]
Year of publishing:	2025
Number of pages:	1 spletni vir (1 datoteka PDF (VII, 63 str.))
PID:	20.500.12556/DKUM-94605
UDC:	004.8.032.26:004.932(043.2)
COBISS.SI-ID:	256081411
Publication date in DKUM:	17.10.2025
Views:	165
Downloads:	31
Metadata:
Categories:	KTFMB - FERI
:	Copy citation

Average score:	(0 votes)
Your score:	Voting is allowed only for logged in users.
Share:

Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:	CC BY-NC-ND 4.0, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

Link:	http://creativecommons.org/licenses/by-nc-nd/4.0/
Description:	The most restrictive Creative Commons license. This only allows people to download and share the work for no commercial gain and for no other purposes.
Licensing start date:	22.08.2025

Secondary language

Language:	English
Title:	Optimization of deep networks for facial expression recognition
Abstract:	As part of this master's thesis, we addressed the complex problem of facial expression recognition using deep neural networks. Current solutions are extremely inefficient when applied to the real world. In the practical part, we used a differential evolution algorithm to optimize the parameters of three well-known deep neural network architectures—DenseNet121, ResNet50, and VGG16—with the aim of achieving higher accuracy. Using the AffectNet training set and transfer learning techniques, we adapted the models for the task of emotion recognition. With differential evolution, we successfully found hyperparameters (learning rate, momentum, weight decay) that give us slightly better results for each model. We increased the accuracy of DenseNet121 by 1.72% (from 52.24% to 53.96%), ResNet by 1.88% (from 51.86% to 53.74%), and VGG16 by 1.59% (from 52.01% to 53.6%). . We also observed how the size of the set affects performance. We found that subsets ranging from 30% to 40% of the total give results that are on average about 2% worse than the results where we used the entire training set.
Keywords:	emotion recognition, convolutional neural networks, learning parameter optimization

Comments

Leave comment

You must log in to leave a comment.

Comments (0)

0 - 0 / 0

There are no comments!

Back