Date of Award

12-2018

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Dr. Ikhlas M. Abdel-Qader

Second Advisor

Dr. Janos Grantner

Third Advisor

Dr. Azim Houshyar

Abstract

Facial recognition is still a challenge in many applications, particularly in surveillance, security systems, and human-computer interactive (HCI) tools. The impact of real-world environmental variations on the performance of any facial recognition system can be significant. These variations may include illumination, facial expressions, poses, disguises (facial hair, glasses or cosmetics) and partial face occlusion. Furthermore, the number of images (samples) needed for facial recognition systems can be very large, much larger than what is typically required by an algorithm, and this is often made worse with expansion of the feature space dimensionality. Truly, the “curse of dimensionality” haunts any real-time implementation of the majority of proposed algorithms. Generally, in all facial recognition applications, high-dimensional facial representations with massive face datasets have caused serious challenges in the cost of implementation.

In general, facial recognition systems are composed of two main stages: a) the feature extraction stage followed by b) the classification stage. In this study, a hybrid feature extraction method to enhance speed and recognition efficiency is proposed based on features obtained using the Histograms of Oriented Gradients (HOG) descriptor and Compressive Sensing (CS). The HOG feature descriptor has the advantage of extracting representative facial feature vectors even with changes in face appearance and is fully capable of handling variations in illumination. CS is used to reduce the density of the resulting HOG facial features, which has a significant impact on improving the computational cost and performance of the system. For 2 the classification stage, the k-Nearest Neighbors (k-NN) algorithm and Probabilistic Neural Network (PNN) classifier are used with this proposed reduced feature space. The results using Face96, Caltech, and AR face datasets demonstrate that this hybrid feature-extraction method could be developed as a complete system for identifying faces even with varying illumination, facial expressions, poses, occlusion, and backgrounds in real-time.

Additionally, a new feature extraction framework using another feature of HOG, which is the ability to vary its parameter values and thus feature dimension, is proposed, allowing for improved feature selection and, consequently, the detection of facial features. The HOG feature extraction mechanisms followed by the CS-based classification stage allows for better recognition rates with a minimum feature dimension and significant reduction in computational time. The results of this framework are presented and compared with those of a PNN-based classification algorithm. Also, using the ORL face, JAFFE face and AR face datasets, this study demonstrates that this method is capable of handling both dimensionality challenges and environmental variations. Random dataset splitting techniques and cross-validation evaluation approach with different k-fold values are also proposed to use in parallel with CS, PNN and k-NN methods for optimum model selection and a complete analysis of system performance.

Using the five national face datasets, the experimental results demonstrate that there are significant improvements in accuracy and performance of the recognition system while combating the dimensionality challenge, face occlusions, facial expression, pose and illumination challenges, memory requirements, and computational complexity.

Access Setting

Dissertation-Campus Only

Restricted to Campus until

12-2019

Share

COinS