Prijenosno učenje za određivanje smjera gledanja

master's thesis

• Title (english): Enhancing transfer learning for head pose estimation
• Author: Josip Matak
• Mentor: Domagoj Jakobović (mentor)
• Committee member: Domagoj Jakobović (predsjednik povjerenstva)
• Committee member: Marko Đurasević (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Odred̄ivanje smjera gledanja korisna je praksa u mnogim industrijama, naročito onim
koje su medicinske, tehničke ili automobilske prirode. Preciznost i brzina izvod̄enja ključne
su stvari koje ove metode čine prilagod̄enima korištenju u praksi.
Rješenje za takvu zadaću nudi sve češće korištena grana strojnog učenja, duboko učenje.
Glavni doprinos ovog rada je tehnika prijenosnog učenja, te nekoliko metoda pretrage najbolje arhitekture neuronske mreže, sagrad̄ene od prije definiranih blokova. Rezultati koji su
izloženi u radu sugestiraju da su značajke izvučene iz metode prepoznavanja lica dovoljne
za ekstrakciju smjera gledanja. Takod̄er, metode optimiranja arhitekture dubokog modela pokazale su unaprijed̄enje brzine izvod̄enja i preciznosti optimiranjem višekriterijske
funkcije cilja. Iako je zadatak koji zahtjeva vremenske i memorijske resurse, automatizirana
pretraga omogućuje usmjeravanje ka drugim smjerovima istraživanja istog problema.

Abstract (english)

Estimating head pose is a useful practice in many industries, especially those of a medical, technical, or automotive nature. Precision and speed of execution are the key things that
make these methods useful in real-world practice.
The solution to this task is offered by the current hot topic in deep learning, a machine
learning subset. The main contribution of this paper is the transfer learning technique and
several different search methods for the best neural network architecture, built from predefined blocks. The results presented in the paper suggest that features extracted from the face
recognition method are sufficient for estimating head pose. Also, deep model architecture
optimization methods have shown improvements in execution speed and precision by optimizing the multi-criterion objective function. Although this is a task that requires time and
memory resources, automated search allows users to change the main focus of research into
different problems in the same domain.

• Keywords: Prijenosno učenje; Duboko učenje; Evolucijsko računarstvo; Estimacija smjera gledanja; Pretraga arhitekture dubokog modela; Heurističke metode
• Keywords (english): Transfer learning; Deep learning; Evolutionary computing; Heads pose estimation; Neural architecture search; Heuristic methods
• Language: english
• URN:NBN: urn:nbn:hr:168:952510
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 13:03:23