[ Chris Fallin ]
This was an undergraduate research project with Professor Flynn, begun in Fall 2007 and running through Spring 2009 (for a total of four semesters), which meandered along a fairly circuitous route through facial recognition and finally arrived at a specific subproblem of the Elastic Bunch Graph Matching (EBGM) facial recognition algorithm. I used the Colorado State Facial Identification Evaluation System as a foundation.
Specifically, I built a system that chose fiducial points automatically, replacing the manually-placed fiducial points that the EBGM system uses to initialize its bunch graph. In theory, the system is completely general and thus not restricted to frontal face images, requiring only two anchor points (nominally, the eye coordinates) in order to align the images. The primary purpose of the project was to prove that the manual, labor-intensive bunch graph training process was unnecessary and that statistical and image-processing methods could provide nearly equivalent matching performance.
The system achieved a final rank-one matching rate of 81% on the FERET face image dataset, nearly equivalent to the 82% rate achieved using the original hand-placed points. Significant potential for further work remains.
This project served as the basis of my undergraduate thesis for the Engineering Honors Program at Notre Dame.