Electronic Theses and Dissertations

Identifier

486

Teddy Salan

Date

2011

Document Type

Thesis

Degree Name

Master of Science

Major

Electrical and Computer Engr

Concentration

Computer Engineering

Committee Chair

Khan M Iftekharuddin

Committee Member

Peter Lau

Committee Member

Aaron L Robinson

Abstract

Face recognition is a widely covered and desirable research field that produced multiple techniques and different approaches. Most of them have severe limitations with pose variations or face rotation. The immediate goal of this thesis is to deal with pose variations by implementing a face recognition system using a Cellular Simultaneous Recurrent Network (CSRN). The CSRN is a novel bio-inspired recurrent neural network that mimics reinforcement learning in the brain. The recognition task is defined as an identification problem on image sequences. The goal is to correctly match a set of unknown pose distorted probe face sequences with a set of known gallery sequences. This system comprises of a pre-processing stage for face and feature extraction and a recognition stage to perform the identification. The face detection algorithm is based on the scale-space method combined with facial structural knowledge. These steps include extraction of key landmark points and motion unit vectors that describe movement of face sequqnces. The identification process applies Eigenface and PCA and reduces each image to a pattern vector used as input for the CSRN. In the training phase the CSRN learns the temporal information contained in image sequences. In the testing phase the network predicts the output pattern and finds similarity with a test input pattern indicating a match or mismatch.Previous applications of a CSRN system in face recognition have shown promise. The first objective of this research is to evaluate those prior implementations of CSRN-based pose invariant face recognition in video images with large scale databases. The publicly available VidTIMIT Audio-Video face dataset provides all the sequences needed for this study. The second objective is to modify a few well know standard face recognition algorithms to handle pose invariant face recognition for appropriate benchmarking with the CSRN. The final objective is to further improve CSRN face recognition by introducing motion units which can be used to capture the direction and intensity of movement of feature points in a rotating face

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.

Download

Share

COinS