Date of Award

12-2014

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Computer Science

Advisor

Woodard, Damon L

Committee Member

Daily , Shaundra

Committee Member

Gilbert , Juan

Committee Member

Hallstrom , Jason

Abstract

As the number of biometric applications increases, the use of non-ideal information such as images which are not strictly controlled, images taken covertly, or images where the main interest is partially occluded, also increases. Face images are a specific example of this. In these non-ideal instances, other information, such as gender and ethnicity, can be determined to narrow the search space and/or improve the recognition results. Some research exists for gender classification using partial-face images, but there is little research involving ethnic classifications on such images. Few datasets have had the ethnic diversity needed and sufficient subjects for each ethnicity to perform this evaluation. Research is also lacking on how gender and ethnicity classifications on partial face are impacted by age. If the extracted gender and ethnicity information is to be integrated into a larger system, some measure of the reliability of the extracted information is needed. This study will provide an analysis of gender and ethnicity classification on large datasets captured by non-researchers under day-to-day operations using texture, color, and shape features extracted from partial-face regions. This analysis will allow for a greater understanding of the limitations of various facial regions for gender and ethnicity classifications. These limitations will guide the integration of automatically extracted partial-face gender and ethnicity information with a biometric face application in order to improve recognition under non-ideal circumstances. Overall, the results from this work showed that reliable gender and ethnic classification can be achieved from partial face images. Different regions of the face hold varying amount of gender and ethnicity information. For machine classification, the upper face regions hold more ethnicity information while the lower face regions hold more gender information. All regions were impacted by age, but the eyes were impacted the most in texture and color. The shape of the nose changed more with respect to age than any of the other regions.

Share

COinS