Date of Award

Fall 12-2015

Degree Type


Degree Name

Doctor of Philosophy (PhD)




Computing Sciences and Computer Engineering

Committee Chair

Zhaoxian Zhou

Committee Chair Department


Committee Member 2

Chaoyang Zhang

Committee Member 2 Department


Committee Member 3

Zheng (Jonathan) Sun

Committee Member 3 Department


Committee Member 4

Zheng Wang

Committee Member 4 Department


Committee Member 5

Ras Pandey

Committee Member 5 Department

Physics and Astronomy


For every face recognition method, the primary goal is to achieve higher recognition accuracy and spend less computational costs. However, as the gallery size increases, especially when one probe image corresponds to only one training image, face recognition becomes more and more challenging. First, a larger gallery size requires more computational costs and memory usage. Meanwhile, that the large gallery sizes degrade the recognition accuracy becomes an even more significant problem to be solved.

A coarse parallel algorithm that equally divides training images and probe images into multiple processors is proposed to deal with the large computational costs and huge memory usage of the Non-Graph Matching (NGM) feature-based method. First, each processor finishes its own training workload and stores the extracted feature information, respectively. And then, each processor simultaneously carries out the matching process for their own probe images by communicating their own stored feature information with each other. Finally, one processor collects the recognition result from the other processors. Due to the well-balanced workload, the speedup increases with the number of processors and thus the efficiency is excellently maintained. Moreover, the memory usage on each processor also evidently reduces as the number of processors increases. In sum, the parallel algorithm simultaneously brings less running time and memory usage for one processor.

To solve the recognition degradation problem, a set of multi-stage matching algorithms that determine the recognition result step-by-step are proposed. Each step picks a small proportion of the best similar candidates for the next step and removes the others. The behavior of picking and removing repeats until the number of remaining candidates is small enough to produce the final recognition result. Three multi-stage matching algorithms— n-ary elimination, divide and conquer, and two-stage hybrid— are introduced to the matching process of traditional face recognition methods, including Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Non-graph Matching (NGM). N-ary elimination accomplishes the multi-stage matching from the global perspective by ranking the similarities and picking the best candidates. Divide and conquer implements the multi-stage matching from the local perspective by dividing the candidates into groups and selecting the best one of each group. For two-stage hybrid, it uses a holistic method to choose a small amount of candidates and then utilizes a feature-based method to find out the final recognition result from them. From the experimental results, three conclusions can be drawn. First, with the multi-stage matching algorithms, higher recognition accuracy can be achieved. Second, the larger the gallery size, the greater the improved accuracy brought by the multi-stage matching algorithms. Finally, the multi-stage matching algorithms achieve little extra computational costs.