Razpoznavanje šarenice z uporabo globokega učenja

Lozej, Juš (author), Peer, Peter (mentor), Štruc, Vitomir (comentor)

Kljub izjemni rasti globokega učenja v zadnjih letih, do sedaj še ni bil razvit globok cevovod za razpoznavo šarenice. V nadaljevanju predstavljamo splošno arhitekturo za razpoznavo šarenice navdihnjeno po strukturi tradicionalnega cevovoda za razpoznavo šarenice. Naš cevovod je zaključena konvolucijska nevronska mreža sestavljena iz dveh visoko-nivojskih gradnikov: segmentacije in razpoznave. Z množenjem izhoda segmentacijskega dela model izloči področja, ki ne pripadajo šarenici in jih poda razpoznavi. Razpoznavni del iz šarenice izlušči značilke, katere uporabimo pri razpoznavi oseb. Naša metoda je na testnih podatkovnih zbirkah dosegla visoke rezultate. Na zbirki Casia-Iris-Thousand je dosegla natančnost prvega ranga 95,12 % in na zbirki SBVPI natančnost 92,33 %. Implementirali smo tudi med-podatkovni model, naučen na vzorcih obeh zbirk, ki je na sklopu zbirk dosegel natančnost 88,53 %. Naša metoda je presegla uspešnost in hitrost tradicionalnega cevovoda. Naš cevovod, kolikor vemo, predstavlja prvo implementacijo globoke nevronske mreže, ki znotraj svoje strukture segmentira področje šarenice in to nato razpozna. Za razliko od trenutnih pristopov naš razpozna osebo na podlagi izvorne nenormalizirane slike očesa.

• Language: Slovenian
• Keywords: Razpoznavanje šarenice, globoko učenje, konvolucijske nevronske mreže, biometrija
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-105260
• Publication date in RUL: 15.11.2018

LOZEJ, Juš, 2018, Razpoznavanje šarenice z uporabo globokega učenja [online]. Master’s thesis. [Accessed 27 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=105260

Secondary language

• Language: English
• Title: Iris recognition using deep learning

Abstract

Despite the large increase of deep learning solutions in recent years, no deep learning iris pipelines have yet been developed. Inspired by conventional iris recognition pipelines, we present our general deep architecture for iris recognition. The presented deep iris pipeline is an end-to-end convolutional neural network consisting of two high-level blocks: segmentation and recognition. The segmentation part is tasked with the generation of binary mask, which corresponds with the surface of the iris. These masks are multiplied with the original iris image and then fed to the recognition part. The recognition part extracts meaningful iris features, which are then used for matching. Our model achieved high results on both testing datasets. On Casia-Iris-Thousand it achieved a Rank-1 accuracy of 95.12% and on SBVPI an accuracy of 92.33%. We also implemented a cross-database model, trained on samples from both dataset, which achieved an accuracy of 88.53%. Our deep pipeline outperformed a conventional iris pipeline in speed and accuracy. As far as we are aware, our pipeline is the rst implementation of an end-to-end deep neural network, which is able to segment and recognize the iris image. As opposed to current deep models, which perform recognition on a pre-normalized iris image, our method uses original iris images.

• Keywords: Iris recognition, deep learning, convolutional neural networks, biometrics