Ear Alignment using Deep Learning

Hrovatič, Anja (author), Peer, Peter (mentor), Emeršič, Žiga (co-mentor)

Ear biometrics identification methods have become very popular in the recent years, especially since the ear presents itself as a reliable modality for recognition with attractive qualities of universality, uniqueness, measurability and permanence. In early ear recognition research, alignment has always been used as a preprocessing step to ensure reliable and robust verification and recognition systems. However, lately the ear recognition research has mostly been oriented towards obtaining better features, omitting the alignment step completely. In our research we tackle the problem of ear alignment by employing deep learning methods. We develop a framework for automatic landmark localization on 2D ears of the "In-the-wild" Ear dataset, employing means of data augmentation to obtain a large-scale dataset with annotated landmarks that is further used to train deep learning architectures. We perform landmark fitting experiments on the ITWE and AWE datasets and obtain results superior to state-of-the-art with the use of two Stack Hourglass Network architecture. Lastly, we employ landmark-based geometric normalization technique to obtain aligned ear images of both datasets and perform recognition experiments on both unaligned and aligned data.

• Language: English
• Keywords: ear biometrics, computer vision, deep learning
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-133721
• COBISS.SI-ID: 91308803
• Publication date in RUL: 10.12.2021

Secondary language

• Language: Slovenian
• Title: Poravnava uhljev z uporabo globokega učenja

Abstract

Raziskave na področju biometrije uhljev so v zadnjih letih postale vse bolj popularne, predvsem ker uhelj predstavlja zanesljivo modalnost za identifikacijo, saj ima privlačne lastnosti, kot so univerzalnost, edinstvenost, merljivost in trajnost. V začetnih obdobjih raziskav na področju prepoznave uhljev za identifikacijo se je postopek poravnave zmeraj uporabil kot eden izmed postopkov predpriprave podatkov za zagotovitev zanesljivih in robustnih sistemov za verifikacijo ter identifikacijo na podlagi uhljev. Zadnje čase pa raziskovalci svojo pozornost usmerjajo predvsem v pridobivanje boljših atributov, s čimer popolnoma izpuščajo postopek poravnave. V sklopu naše raziskave se ukvarjamo s problemom poravnave uhljev z uporabo metod globokega učenja. Razvijemo ogrodje za avtomatsko razpoznavo značilk na 2D slikah uhljev podatkovne množice ITWE. Prav tako koristimo postopke obogatitve podatkov, da povečamo velikost podatkovne množice ITWE. Le-to nato uporabimo za učenje globokih arhitektur. Izvedemo eksperimente na dveh podatkovnih nizih, in sicer na ITWE in AWE podatkovnih bazah, ter poročamo o vrhunskih rezultatih v primerjavi z najboljšimi na področju poravnave uhljev z uporabo arhitekture dvonivojske mreže peščene ure. Nazadnje izvedemo geometrično poravnavo uhljev na podlagi značilk in na le-teh izvedemo identifikacijske eksperimente.

• Keywords: biometrija uhljev, računalniški vid, globoko učenje