Date of Award

12-2020

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering and Sciences

First Advisor

Michael King

Second Advisor

Kevin Bowyer

Third Advisor

Munevver Mine Subasi

Fourth Advisor

Philip Bernhard

Abstract

Researchers seek methods to increase the accuracy and efficiency in identifying an individual using facial biometric systems. Factors like skin color, which may affect the accuracy of facial recognition, need to be investigated further. To analyze the impact of race with respect to skin color of face on biometric systems, we focused on generating a dataset with uniformly distributed images of different skin tones while preserving identities. Deep learning neural network architectures like the generative adversarial network (GAN) focus on modifying only certain features of the face like the color of the skin. In our experimental approach, we implemented a cycle GAN to synthesize multiple images of individuals with varying color of their skin based on the Fitzpatrick scale. The Fitzpatrick scale defines six levels of skin tone, with FP-1 representing a lighter and FP-6 a darker shade of color. The resulting GAN receives an image of a person with a skin tone labeled FP-1 as its input and synthesizes five additional images to show what that person would look like with skin-tone ratings of FP-2 to FP-6. A GAN was trained to receive an image of a person with a skin-tone rating of FP-6 and subsequently to produce images corresponding to skin tone ratings of FP-5 to FP-1. The Arcface matcher was used to measure the similarity between the original images of a person and those produced by the GAN. The results of the analysis indicate a drop in the similarity scores when skin tones change from light to dark and vice versa. In future work, we intend to train a facial recognition algorithm to evaluate the impact of bias relative to skin tone with uniformly generated improved version of skin color images of the same individual.

Comments

Copyright held by author

Share

COinS