Biometric refinement of datasets for facial age estimation

doctoral thesis

• Title (english): Biometrijsko unapređenje skupova podataka za procjenu dobi iz slika lica
• Author: Krešimir Bešenić
• Mentor: Igor Sunday Pandžić (mentor)
• Mentor: Jörgen Ahlberg https://orcid.org/0000-0002-6763-5487 (komentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

This thesis explores the intricate task of facial age estimation with a focus on the role of data in the design of machine-learning-based systems. Supervised machine learning algorithms rely on large amounts of labeled data to learn a mapping between the input data (e.g., facial images or videos) and associated labels (e.g., age or gender). Due to difficulties related to biometric data collection, facial image datasets with biometric trait labels are scarce and frequently limited in terms of size and sample diversity. Web-scraping approaches for automatic data collection can produce large amounts of heterogeneous but weakly labeled and noisy data. The portion of erroneous samples in such data can be automatically reduced by the proposed unsupervised biometric data filtering method. The problem of data scarcity is further addressed by designing a new public facial image dataset. The proposed derived dataset is thoroughly compared to various state-of-the-art datasets with very favorable results and interesting insights about the significance of data. Video sequences can capture many spatiotemporal dynamic cues related to age progression. However, video-based age estimation is an understudied research field. This may be caused by a lack of well-defined and publicly accessible video benchmark protocols, as well as the absence of video-oriented training data. The former issue was addressed by proposing a carefully designed video age estimation benchmark protocol and making it publicly available. The latter issue was addressed by designing a video-specific age estimation method that leverages pseudo-labeling and semi-supervised learning. The experimental results indicate that using video data unlocks great potential in the age estimation research field.

Abstract (croatian)

Ovaj rad istražuje složen zadatak procjene dobi iz slika lica s naglaskom na ulogu podataka u dizajnu sustava temeljenih na strojnom učenju. Algoritmi nadziranog strojnog učenja oslanjaju se na velike količine označenih podataka kako bi naučili mapiranje između ulaznih podataka (npr. slika ili videozapisa lica) i pripadajućih oznaka (npr. dobi ili spola). Zbog poteškoća vezanih uz prikupljanje biometrijskih podataka, skupovi slika lica s biometrijskim oznakama su rijetki i često ograničeni u pogledu veličine i raznolikosti uzoraka. Metodama za prikupljanje podataka iz internetskih izvora mogu se skupiti velike količine heterogenih, ali nepouzdano označenih podataka. Udio pogrešnih uzoraka u takvim skupovima podataka može se automatski smanjiti predloženom metodom za nenadzirano biometrijsko filtriranje podataka. Problem nedostatka podataka za učenje dodatno je adresiran dizajniranjem nove javno dostupne baze slika lica. Predloženi izvedeni skup podataka temeljito je uspoređen s trenutno najboljim javno dostupnim skupovima podataka s vrlo povoljnim rezultatima i zanimljivim uvidima o značaju podataka. Videozapisi mogu sadržavati mnoge prostorno-vremenske dinamičke značajke relevantne za procjenu dobi, no procjena dobi iz video podataka je slabo istraženo znanstveno polje. Potencijalni uzrok za to je nedostatak dobro definiranih i javno dostupnih evaluacijskih protokola, kao i nedostatak označenih video podataka za učenje. Prvi od navedenih problema riješen je dizajniranjem javno dostupnog protokola za procjenu dobi iz video podataka. Drugom problemu pristupilo se dizajniranjem metode za procjenu dobi iz video podataka koja koristi pseudo-označavanje i djelomično nadzirano učenje. Eksperimentalni rezultati pokazuju da korištenje video podataka otkriva veliki potencijal u području istraživanja procjene dobi.

• Keywords: face analysis; age estimation; deep learning; data filtering; dataset design; semi-supervised learning; temporal models
• Keywords (croatian): analiza lica; procjena dobi; duboko ucenje; filtriranje podataka; dizajn skupova podataka; djelomično nadzirano učenje; temporalni modeli
• Language: english
• URN:NBN: urn:nbn:hr:168:606173
• Promotion: 2024
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: 135 str.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-09-03 12:44:49