Face Recognition Using Foveal Vision

Abstract

Data from human subjects recorded by an eyetracker while they are learning new faces shows a high degree of similarity in saccadic eye movements over a face. Such experiments suggest face recognition can be modeled as a sequential process, with each fixation providing observations using both foveal and parafoveal information. We describe a sequential model of face recognition that is incremental and scalable to large face images. Two approaches to implementing an artificial fovea are described, which transform a constant resolution image into a variable resolution image with acute resolution in the fovea, and an exponential decrease in resolution towards the periphery. For each individual in a database of faces, a hidden-Markov model (HMM) classifier is learned, where the observation sequences necessary to learn the HMMs are generated by fixating on different regions of a face. Detailed experimental results are provided which show the two foveal HMM classifiers outperform a more traditional HMM classifier built by moving a horizontal window from top to bottom on a highly subsampled face image.