prikaz prve stranice dokumenta Raspoznavanje objekata uporabom umjetnih neuronskih mreža
No public access
undergraduate thesis
Raspoznavanje objekata uporabom umjetnih neuronskih mreža
Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing, 2020. urn:nbn:hr:168:097422
Arambašić, Lucija
University of Zagreb
Faculty of Electrical Engineering and Computing

Institutional repository: FER Repository

Cite this document

Arambašić, L. (2020). Raspoznavanje objekata uporabom umjetnih neuronskih mreža (Undergraduate thesis). Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing. Retrieved from https://urn.nsk.hr/urn:nbn:hr:168:097422

Arambašić, Lucija. "Raspoznavanje objekata uporabom umjetnih neuronskih mreža." Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2020. https://urn.nsk.hr/urn:nbn:hr:168:097422

Arambašić, Lucija. "Raspoznavanje objekata uporabom umjetnih neuronskih mreža." Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2020. https://urn.nsk.hr/urn:nbn:hr:168:097422

Arambašić, L. (2020). 'Raspoznavanje objekata uporabom umjetnih neuronskih mreža', Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, accessed 13 July 2024, https://urn.nsk.hr/urn:nbn:hr:168:097422

Arambašić L. Raspoznavanje objekata uporabom umjetnih neuronskih mreža [Undergraduate thesis]. Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing; 2020 [cited 2024 July 13] Available at: https://urn.nsk.hr/urn:nbn:hr:168:097422

L. Arambašić, "Raspoznavanje objekata uporabom umjetnih neuronskih mreža", Undergraduate thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, 2020. Available at: https://urn.nsk.hr/urn:nbn:hr:168:097422

Metadata
TitleRaspoznavanje objekata uporabom umjetnih neuronskih mreža
Title (english)Object Recognition Using Artificial Neural Networks
AuthorLucija Arambašić
MentorZoran Kalafatić (mentor)
Committee memberZoran Kalafatić (predsjednik povjerenstva)
Committee memberJosip Šarić (član povjerenstva)
Committee memberPetra Bevandić (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb
Defense date and country2020-09-02, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Abstract
Umjetne neuronske mreže predstavljaju pokušaj imitiranja građe ljudskog mozga pri rješavanju problema koje ljudi s lakoćom rješavaju. U sklopu završnog rada proučene su duboke neuronske mreže te njihova primjena u području računalnog vida, konkretnije u raspoznavanju objekata na slikama. Proučene se različite arhitekture modela te je formiran jednostavan model duboke mreže. Nad tim modelom provedeni su postupci regularizacije s ciljem poboljšavanja sposobnost generalizacije modela. Model je postizao dobre rezultate nad skupom za učenje, ali nakon određenog broja epoha mu se točnost nad skupom za provjeru počela smanjivati. To je bio očit znak da se model počeo specijalizirati za podatke iz skupa za učenje. Kako bi se to umanjilo implementirane su 3 tehnike regularizacije: generiranje podataka, rano zaustavljanje te isključivanje čvorova. Posljedice su odmah bile vidljive, učinkovitost modela nad skupom za učenje se malo smanjila, ali učinkovitost nad skupom za provjeru se poboljšala. Rano zaustavljanje je spriječilo prenaučenost modela jer se proces učenja prekine čim greška na skupu za provjeru počne rasti. Generiranje podataka i isključivanje čvorova su povećali dotad najbolju točnost nad skupom za provjeru. Nakon učenja model s podešenim parametrima ispitan je na skupu za ispitivanje, nad kojim je ostvario točnost od 93,39\%, što je poprilično dobar rezultat.
Abstract (english)
Artificial neural networks represent an imitation of the human brain used for solving problems which humans solve with ease. As part of the bachelor's thesis deep neural networks and their usage for object classification were studied. After researching several different deep neural network model architectures a simple model was formed. Regularization techniques were conducted on the model in an attempt to increase its generalization ability. Model performed well on training set, but after a certain number of epochs its accuracy on the validation set begun to decrease. That was the clear sign the model started to overfit the data in the training set. To reduce this 3 regularization techniques were implemented: data augmentation, early stopping and dropout. The effects were immediately visible, the model's accuracy on the training set decreased, but its accuracy on the validation set increased. Early stopping prevented the model from overfitting because the training process stops the moment the loss on validation set starts to increase. Data augmentation and dropout increased the so far highest accuracy on the validation set. After the training process the model was tested on the test set, and it achieved the accuracy of 93,39\%, which is a pretty good result.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:168:097422
Study programmeTitle: Computing
Study programme type: university
Study level: undergraduate
Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsClosed access
Terms of useLicence In copyright icon
Public note
Created on2023-01-14 08:20:05