izpis_h1_title_alt

Analiza primernosti globokih nevronskih omrežij za razpoznavanje šarenice
ID KROFIČ, MATEJ (Author), ID Štruc, Vitomir (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (2,22 MB)
MD5: C07D53570D4A3D7B13466693410851F4
PID: 20.500.12556/rul/c212e8e2-6de4-4d4a-9d39-9be5896b4f0d

Abstract
Tehnologija in znanje, ki sta bila do nedavno uveljavljena le v laboratorijih in v visokih krogih znanosti, sedaj postajata vedno bolj dostopna širšemu krogu ljudi. Vse bolj zmogljivi osebni računalniki dovoljujejo, da se lahko sistemi, katerih namen je iskanje unikatnih značilnosti področja šarenice, implementirajo kjerkoli, ne da bi razpolagali z visokotehnološki ali drago opremi. Tovrstni sistemi so prisotni tudi na višjih nivojih varnosti; od Združenih Arabskih Emiratov, ki ta način preverjanja identitete uporabljajo več kot desetletje; Indije, ki področje šarenice uporablja v sklopu programa Unique ID; do številnih letališč, ki uporabljajo biometrične metode identifikacije za pohitritev procesa preverjanja pogojev za prestop meje. Glede na širino področja uporabe in potrebe po vedno hitrejših in zanesljivejših algoritmih, je iskanje alternativnih metod nujno. Cilj tega diplomskega dela je raziskati alternativne metode razpoznavanja področja šarenice in uspešnost primerjati s klasičnimi metodami. Kot alternativo smo izbrali metodo pridobitve vektorjev značilk s pomočjo vnaprej naučenega globokega konvolucijskega nevronskega omrežja GoogLeNet. Za vhodne slikovne datoteke smo uporabili slike iz prosto dostopne baze podatkov CASIA V4, ki jo je izdala Kitajska akademija znanosti (Chinese Academy of Sciences). Zbirka ima šest različnih pod zbirk s poudarkom na različnih lastnostih. Za namen diplomskega dela smo izbrali štiri pod zbirke, obdelali smo jih s tremi različnimi postopki. Prvi je klasičen postopek pridobivanja vektorjev značilk; drugi je hibriden, kjer je vhodna slikovna datoteka v nevronsko omrežje segmentirana in normalizirana slika šarenice; tretji je direkten, vhodna datoteka pa je v globoko nevronsko omrežje neobdelana slika. Rezultat raziskave je poglobljena analiza in primerjava vseh treh predstavljenih metod.

Language:Slovenian
Keywords:biometrija, šarenica, globoka nevronska omrežja
Work type:Undergraduate thesis
Organization:FE - Faculty of Electrical Engineering
Year:2016
PID:20.500.12556/RUL-85008 This link opens in a new window
Publication date in RUL:09.09.2016
Views:1462
Downloads:381
Metadata:XML RDF-CHPDL DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:English
Title:A feasibility study of deep neural networks for iris recognition
Abstract:
The knowledge and technologies that used to be limited to laboratories and academia, are becoming increasingly available to everyone. The growing capabilities of personal computers allow for the implementation of systems that recognize unique iris features anywhere, without the need for high-end or expensive equipment. These types of iris recognition systems are being used in environments with even the highest security demands. The United Arab Emirates have been using this kind of person identification for more than a decade, India is using iris recognition as part of its Unique ID program and a number of airport are using biometric identification methods to speed up border security check. Given the wide range of uses and the need for faster, more reliable and secure algorithms, the search for alternative methods is necessary and vital. The goal of this thesis is to research alternative methods of iris recognition and to determine how successful they are compared to conventional methods. As an alternative method we chose computing feature vectors using a pre-trained deep convolutional neural network called GoogLeNet. The input image files used for the experiments were from a freely available image database CASIA V4, published by the Chinese Academy of Sciences. The collection has six different subsets with different attributes or characteristics. For the purpose of the thesis we have chosen only four subsets and ran them through three different procedures. The first is the conventional method computing feature vectors; the second is the hybrid method, in which the input image in a neural network is a segmented and normalized iris image; the third is the direct method in which the input image file in the deep neural network is an unedited iris image. The research result is a deep analysis and comparison of the three presented methods.

Keywords:biometrics, iris, deep neural networks

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back