Analysis of color effect on the face recognition with spatial resolution constraints

Abstract

Human face is the most common and natural biometric signature to distinguish different identities. However, there remain many restrictions in automatic face recognition (FR) due to illumination, pose, aging variations, small size and low image quality. Particularly, face resolution is a significant constraint to some FR applications (e.g., surveillance-related and access control systems) where various resolutions could be obtained due to different camera capture conditions. Under such FR environments, it is important to find face cues robust to resolution variations. Color feature is generally less vulnerable to image degradation and variation in resolution relative to grayscale. We investigate the effect of color on FR with resolution variations in well-known appearance-based method, PCA and LDA. In FR applications like surveillance being confined to resolution limitations, the practical issue is that the resolution of registered face is different from that used for verification or identification. To deal with this problem, we present an estimation of feature subspace that optimally represents lower resolution faces from given feature subspace pre-constructed with high-resolution faces. Also, color is applied to our proposed subspace estimation method to observe the effect on performance with respect to resolution changes. The theoretical analysis and extensive experiments are given to verify color role in FR constrained with low face resolution.