Date on Master's Thesis/Doctoral Dissertation


Document Type

Master's Thesis

Degree Name

M. Eng.


Electrical and Computer Engineering

Committee Chair

Graham, James

Committee Co-Chair (if applicable)

Ralston, Patricia

Committee Member

Ralston, Patricia

Committee Member

Farag, Aly


Research in face recognition deals with problems related to Age, Pose, Illumination and Expression (A-PIE), and seeks approaches that are invariant to these factors. Video images add a temporal aspect to the image acquisition process. Another degree of complexity, above and beyond A-PIE recognition, occurs when multiple pieces of information are known about people, which may be distorted, partially occluded, or disguised, and when the imaging conditions are totally unorthodox! A-PIE recognition in these circumstances becomes really “wild” and therefore, Face Recognition in the Wild has emerged as a field of research in the past few years. Its main purpose is to challenge constrained approaches of automatic face recognition, emulating some of the virtues of the Human Visual System (HVS) which is very tolerant to age, occlusion and distortions in the imaging process. HVS also integrates information about individuals and adds contexts together to recognize people within an activity or behavior. Machine vision has a very long road to emulate HVS, but face recognition in the wild, using the computer, is a road to perform face recognition in that path. In this thesis, Face Recognition in the Wild is defined as unconstrained face recognition under A-PIE+; the (+) connotes any alterations to the design scenario of the face recognition system. This thesis evaluates the Biometric Optical Surveillance System (BOSS) developed at the CVIP Lab, using low resolution imaging sensors. Specifically, the thesis tests the BOSS using cell phone cameras, and examines the potential of facial biometrics on smart portable devices like iPhone, iPads, and Tablets. For quantitative evaluation, the thesis focused on a specific testing scenario of BOSS software using iPhone 4 cell phones and a laptop. Testing was carried out indoor, at the CVIP Lab, using 21 subjects at distances of 5, 10 and 15 feet, with three poses, two expressions and two illumination levels. The three steps (detection, representation and matching) of the BOSS system were tested in this imaging scenario. False positives in facial detection increased with distances and with pose angles above ± 15°. The overall identification rate (face detection at confidence levels above 80%) also degraded with distances, pose, and expressions. The indoor lighting added challenges also, by inducing shadows which affected the image quality and the overall performance of the system. While this limited number of subjects and somewhat constrained imaging environment does not fully support a “wild” imaging scenario, it did provide a deep insight on the issues with automatic face recognition. The recognition rate curves demonstrate the limits of low-resolution cameras for face recognition at a distance (FRAD), yet it also provides a plausible defense for possible A-PIE face recognition on portable devices.