Date on Master's Thesis/Doctoral Dissertation
Electrical and Computer Engineering
Electrical Engineering, PhD
Farag, Aly A.
Committee Co-Chair (if applicable)
Cohn, Robert W.
Yampolskiy, Roman Vladimirovich
Human face recognition (Computer science); Facial expression--Computer simulation; Face--Computer simulation
Face understanding is considered one of the most important topics in computer vision field since the face is a rich source of information in social interaction. Not only does the face provide information about the identity of people, but also of their membership in broad demographic categories (including sex, race, and age), and about their current emotional state. Facial landmarks extraction is the corner stone in the success of different facial analyses and understanding applications. In this dissertation, a novel facial modeling is designed for facial landmarks detection in unconstrained real life environment from different image modalities including infra-red and visible images. In the proposed facial landmarks detector, a part based model is incorporated with holistic face information. In the part based model, the face is modeled by the appearance of different face part(e.g., right eye, left eye, left eyebrow, nose, mouth) and their geometric relation. The appearance is described by a novel feature referred to as pixel difference feature. This representation is three times faster than the state-of-art in feature representation. On the other hand, to model the geometric relation between the face parts, the complex Bingham distribution is adapted from the statistical community into computer vision for modeling the geometric relationship between the facial elements. The global information is incorporated with the local part model using a regression model. The model results outperform the state-of-art in detecting facial landmarks. The proposed facial landmark detector is tested in two computer vision problems: boosting the performance of face detectors by rejecting pseudo faces and camera steering in multi-camera network. To highlight the applicability of the proposed model for different image modalities, it has been studied in two face understanding applications which are face recognition from visible images and physiological measurements for autistic individuals from thermal images. Recognizing identities from faces under different poses, expressions and lighting conditions from a complex background is an still unsolved problem even with accurate detection of landmark. Therefore, a learning similarity measure is proposed. The proposed measure responds only to the difference in identities and filter illuminations and pose variations. similarity measure makes use of statistical inference in the image plane. Additionally, the pose challenge is tackled by two new approaches: assigning different weights for different face part based on their visibility in image plane at different pose angles and synthesizing virtual facial images for each subject at different poses from single frontal image. The proposed framework is demonstrated to be competitive with top performing state-of-art methods which is evaluated on standard benchmarks in face recognition in the wild. The other framework for the face understanding application, which is a physiological measures for autistic individual from infra-red images. In this framework, accurate detecting and tracking Superficial Temporal Arteria (STA) while the subject is moving, playing, and interacting in social communication is a must. It is very challenging to track and detect STA since the appearance of the STA region changes over time and it is not discriminative enough from other areas in face region. A novel concept in detection, called supporter collaboration, is introduced. In support collaboration, the STA is detected and tracked with the help of face landmarks and geometric constraint. This research advanced the field of the emotion recognition.
Mostafa, Eslam AbdelFattah, "Face modeling for face recognition in the wild." (2015). Electronic Theses and Dissertations. Paper 2068.