Prepoznavanje šarenice s pomočjo nevronskih mrež

Polanc, Uroš (author), Batagelj, Borut (mentor)

Naloga obravnava pristop prepoznavanja oseb na podlagi šarenice z nevronskimi mrežami. Ideja je, da na sliki očesa pravilno detektiramo območje šarenice, s katerega nato s primernimi metodami pridobimo tako imenovan vektor značilk. Vektor značilk predstavlja kratek in unikaten opis posamezne slike. Za nevronske mreže smo uporabili klasične nevronske mreže, ki smo jim kot vhod podali vektorje značilk. Na koncu smo preizkusili še konvolucijske nevronske mreže, kjer smo kot vhod podali originalno sliko. Pri klasičnih nevronskih mrežah smo testirali večje število kombinacij metod izboljšave slike, metod izbire značilk ter nevronskih mrež. Izkazalo se je, da mreže za prepoznavanje vzorcev v kombinaciji z Gaborjevimi filtri dosegajo točnost 95,7 procenta. Pri konvolucijskih nevronskih mrežah pa se je najbolje izkazala mreža ResNet50 s točnostjo 96,4 procenta.

• Language: Slovenian
• Keywords: računalniški vid, globoko učenje, nevronske mreže, konvolucijske nevronske mreže, segmentacija šarenice, prepoznavanje šarenice
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-114423
• COBISS.SI-ID: 1538546115
• Publication date in RUL: 27.02.2020

POLANC, Uroš, 2020, Prepoznavanje šarenice s pomočjo nevronskih mrež [online]. Bachelor’s thesis. [Accessed 24 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=114423

Secondary language

• Language: English
• Title: Iris recognition using artificial neural networks

Abstract

The thesis deals with the approach of iris recognition using neural networks. The idea is to correctly detect the iris region from the image of the eye, from which, using suitable algorithms and methods, we then obtain the so-called feature vector. The feature vector represents a compact and unique description of each image, which is then passed to different neural networks. For the neural networks, we use classical neural networks, which are given feature vectors as input. In the end, we also test the convolutional neural networks where the original image is given as input. For classical neural networks, we tested a large number of combinations of image enhancement methods, feature extraction methods and neural networks. Pattern recognition network, in combination with Gabor filters, has been shown to achieve the best accuracy of 95.7 percent. Meanwhile, for convolutional neural networks, the ResNet50 network performed best with an accuracy of 96.4 percent.

• Keywords: computer vision, deep learning, neural network, convolutional neural network, iris segmentation, iris recognition