Feature Subspace Estimation for Resolving Dimensionality Mismatch in Practical Color-Based Face Recognition Systems

Abstract

Recently, considerable research efforts have been devoted to effective utilization of facial color information for improved recognition performance. Of all color-based face recognition (FR) methods, the most widely used approach is a color FR method using input-level fusion. In this method, augmented input vectors of the color images are first generated by concatenating different color components (including both luminance and chrominance information) by column order at the input level and feature subspace is then trained with a set of augmented input vectors. However, in practical applications, a testing image could be captured as a grayscale image, rather than as a color image, mainly caused by different, heterogeneous image acquisition environment. A grayscale testing image causes so-called dimensionality mismatch between the trained feature subspace and testing input vector. Disparity in dimensionality negatively impacts the reliable FR performance and even imposes a significant restriction on carrying out FR operations in practical color FR systems. To resolve the dimensionality mismatch, we propose a novel approach to estimate new feature subspace, suitable for recognizing a grayscale testing image. In particular, new feature subspace is estimated from a given feature subspace created using color training images. The effectiveness of proposed solution has been successfully tested on four public face databases (DBs) such as CMU, FERET, XM2VTSDB, and ORL DBs. Extensive and comparative experiments showed that the proposed solution works well for resolving dimensionality mismatch of importance in real-life color FR systems.