Light-weight deep models for sclera recognition

Vitek, Matej (author), Peer, Peter (mentor), Štruc, Vitomir (comentor)

Sclera recognition is a subfield within biometric recognition technology that focuses on identifying individuals based on the vascular structures in the sclera, i.e. the white part of the eye. Most existing solutions for sclera recognition are based either on hand-crafted methods from the field of computer vision, which perform suboptimally, or on deep convolutional networks, which require powerful hardware to run efficiently. However, biometric systems are increasingly being deployed on smartphones, head-mounted displays, and edge devices, which require light-weight models, i.e. simple computational models capable of running well on weaker hardware. As such, in our thesis (i) we propose the novel method IPAD, which decreases the number of parameters and operations in a deep network, and using IPAD we develop a light-weight model for sclera segmentation, and (ii) we develop the light-weight GazeNet network, based on the SqueezeNet architecture and trained via multi-task learning, which we use as our sclera vessel feature extractor. The results of our extensive experimental analysis affirm the superiority of deep convolutional networks over classical hand-crafted methods. On the other hand, our analysis of the models developed with the IPAD method demonstrates that the networks commonly relied on in the literature can be significantly reduced in terms of their spatial and computational requirements, without a significant decrease in accuracy -- in fact, in certain cases, simplifying the models even enhances their accuracy. Even light-weight deep networks require a significant amount of training data to achieve high-quality performance. We note that, while iris datasets are plentiful, there is a considerable lack of sclera-focused datasets. Thus, as part of the aforementioned contributions, we introduce MOBIUS, the first publicly available mobile-camera-acquired dataset intended primarily for sclera segmentation, although it can be used for iris and periocular biometrics as well. Finally, since biometric systems have been shown to exhibit bias in various biometric fields, we (iii) propose a novel methodology for bias evaluation based on two novel metrics: FSD and CGD. Using the proposed methodology, we study the bias of contemporary sclera segmentation solutions and show that even in sclera biometrics a certain amount of demographic bias is present in existing solutions.

• Language: English
• Keywords: biometrics, identity recognition, lightweight deep neural networks, ocular biometrics, sclera segmentation, sclera recognition
• Work type: Doctoral dissertation
• Typology: 2.08 - Doctoral Dissertation
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-158640
• COBISS.SI-ID: 200362499
• Publication date in RUL: 18.06.2024

Secondary language

• Language: Slovenian
• Title: Lahki globoki modeli za prepoznavo beločnice

Abstract

Razpoznavanje beločnice je področje biometrije, ki se ukvarja s prepoznavo identitete na podlagi žilnih struktur v beločnici, torej v belem delu očesa. Večina pristopov za razpoznavanje beločnice se zanaša ali na klasične ročno zasnovane metode računalniškega vida, ki dosegajo slabše rezultate, ali pa na globoke konvolucijske mreže, ki za dobro delovanje potrebujejo močno strojno opremo. Biometrični sistemi pa se vse bolj nameščajo na pametne telefone, naglavne zaslone in podobne naprave, za katere potrebujemo lahke modele, torej preproste računske modele, ki jih je moč poganjati na manj zmogljivi strojni opremi. V doktorskem delu zato: (i) razvijemo metodo IPAD, s katero zmanjšamo število parametrov in operacij v globokih modelih in z njo razvijemo lahki model za segmentacijo beločnice in (ii) razvijemo lahki model GazeNet, ki temelji na arhitekturi SqueezeNet in je učen s hkratnim učenjem, uporabimo pa ga za luščenje značilk iz žilnih struktur beločnice. Na podlagi obsežnega eksperimentalnega ovrednotenja v delu najprej potrdimo, da globoke nevronske mreže dosegajo občutno boljše rezultate na vseh relevantnih nalogah. Po drugi strani pa z analizo metode IPAD pokažemo, da lahko mreže iz literature znatno zmanjšamo ter kljub temu ohranimo visok nivo natančnosti -- delovanje modela se v nekaterih primerih celo izboljša, ko ga tako poenostavimo. Lahki globoki modeli pa za dobro delovanje še vedno potrebujejo veliko količino učnih podatkov, podatkovnih množic namenjenih biometriji beločnice pa primanjkuje. Zato kot del teh doprinosov sestavimo tudi MOBIUS, ki je prva javno dostopna podatkovna množica primarno namenjena biometriji beločnice s slikami zajetimi s fotoaparatom pametnih telefonov, hkrati pa jo je mogoče uporabiti tudi za raziskave na področju šarenice in periokularne regije. Ker pa so se biometrični sistemi na različnih področjih izkazali za pristranske, v doktorskem delu (iii) predlagamo novo evalvacijsko metodologijo za ocenjevanje pristranskosti biometričnih sistemov na podlagi dveh novih mer: FSD in CGD. V skladu s predlagano metodologijo izvedemo študijo pristranskosti sodobnih pristopov za segmentacijo beločnice in pokažemo, da se tudi na področju beločnice pojavi pristransko delovanje glede na določene demografske faktorje.

• Keywords: biometrija, prepoznava identitete, lahke globoke nevronske mreže, očesna biometrija, segmentacija beločnice, prepoznava beločnice