An Empirical Study of Image Resolution and Pose on Automatic Face Recognition

Abstract

Face image resolution and pose are two important factors that severely degrade the recognition ability. This paper presents a comparison of (i) the Wavelet Transform, (ii) the 2DPCA, (iii) the AdaBoost-LDA, and (iv) Fisherfaces based face recognition algorithms. Simulation results on the Multi-PIE database show that the 2DPCA face recognition algorithm can be reliably used for extremely low face image resolution of 15×15 pixels and from frontal (0°) to +35° of pose variation in near-real time. Whereas for high face image resolution of 40×40 pixels and up to 251×231 pixels, the Fisherfaces yields high accuracy across four different pose variation at the cost of much higher computation. Moreover, the recognition rate of the AdaBoost-LDA is unaffected by the image resolution from 251×231 down to 15×15 pixels. In addition, time cost comparison is also shown.