Author ORCID Identifier

https://orcid.org/0000-0001-5378-4775

Semester

Spring

Date of Graduation

2023

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Jeremy Dawson

Committee Member

Donald Adjeroh

Committee Member

Katerina Goseva-Popstojanova

Abstract

Biometric technology is a rapidly evolving field with applications that range from access to devices to border crossing and entry/exit processes. Large-scale applications to collect biometric data, such as border crossings result in multimodal biometric databases containing thousands of identities. However, due to human operator error, these databases often contain many instances of image labeling and classification; this is due to the lack of training and throughput pressure that comes with human error. Multiple entries from the same individual may be assigned to a different identity. Rolled fingerprints may be labeled as flat images, a face image entered into a fingerprint field or images entered in incorrect orientation (such as rotated face images, left or right iris, etc.) are common errors found large database records. Ultimately, these enrollment errors make it impossible to identify that individual upon subsequent identification encounters. Sorting through hundreds of images to check for classification errors is a tedious and time-consuming task, especially when several biometric databases are combined. Our goal is to correctly identify misclassified fingerprints using controlled embeddings and thresholds. This work provides a new perspective on image sorting as it focuses not on the traditional aspects of increasing accuracy metrics but provides a look into multiple factors through various embeddings and thresholds to provide a tool that can be used to scour large datasets with ease to provide what percentage of the images need manual correction. The proposed network provides various metric scores which allowed for analysis on the most effective embedding and thresholds to use, resulting in a proof-of-concept to be used for practical purposes in the real world.

Download

Share

COinS