Show simple item record

dc.contributor.advisorKing, Michael C.
dc.contributor.authorVangara, Kushal
dc.date.accessioned2018-03-07T20:34:46Z
dc.date.available2018-03-07T20:34:46Z
dc.date.created2018-05
dc.date.issued2018-01
dc.date.submittedMay 2018
dc.identifier.urihttps://repository.lib.fit.edu
dc.identifier.urihttp://hdl.handle.net/11141/2335
dc.descriptionThesis (M.S.) - Florida Institute of Technology, 2018en_US
dc.description.abstractBiometrics technologies are designed for recognition of the unique physiological and behavioral features of an individual for identification and verification applications. There have been noteworthy advances in this field to recognize individuals based on their biometric trait(s) for authentication and verification applications. Despite these advances, there are many challenging issues which impede the potential of biometric systems and therefore limit the systems performance. Traditional biometric system design involves a selection of handcrafted features for recognition tasks which are not efficient as the size of the system scales up. Face and iris biometrics are reliable for many identification and verification applications provided the quality of template extracted is not poor. Periocular recognition has the potential to enhance performance in unconstrained environments. Studies show that periocular biometrics are effective in the cases where face and iris systems fail. There have been significant technical advances in the areas of machine learning and deep learning mainly because of the availability of large datasets and accelerated computing power from graphical processing units (GPU’s). These techniques have an enormous potential especially for the task of image recognition. These methods have state of the art design and have shown high performance on large scale databases. We investigate these techniques by applying them to evaluate performance of face and periocular recognition. But due to the limited availability of periocular datasets, it is challenging to apply these methods for periocular recognition. Hence, we consider the Transfer Learning approach. Transfer Learning, is a research technique in machine learning that focuses on storing knowledge gained while solving one problem and applying it to a different related problem. This work demonstrates face recognition using convolutional neural network (CNN) features and transferring the knowledge of self-learned hierarchical features for the task of periocular recognition. We have used the NDNIVL Dataset consisting of 22,264 near IR images and 574 subjects to evaluate the performance of deep convolutional networks for biometric applications. Our experiments attained an accuracy of 98.93% on face recognition using deep learning method and the results are parallel on comparing with the iris recognition performance using a commercial SDK. By using the transfer learning approach, for periocular recognition we achieved an accuracy of 90.79%. Our results show the importance of periocular region as it has significant discriminative features that can be used in fusion with face or iris for large scale applications involving national security and forensics.en_US
dc.format.mimetypeapplication/pdf
dc.language.isoen_USen_US
dc.rightsCopyright held by author.en_US
dc.titleTransfer Learning with Convolutional Neural Networks Applied to Periocular Biometricsen_US
dc.typeThesisen_US
dc.date.updated2018-03-02T20:36:48Z
thesis.degree.nameMaster of Science in Information Assurance and Cyber Securityen_US
thesis.degree.levelMastersen_US
thesis.degree.disciplineInformation Assurance & Cyber Securityen_US
thesis.degree.departmentComputingen_US
thesis.degree.grantorFlorida Institute of Technologyen_US
dc.type.materialtext


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record