Face recognition through mismatch driven representations of the face

Abstract

Performance of face verification systems can be adversely affected by a number of different mismatches (e.g. illumination, expression, alignment, etc.) between gallery and probe images. In this paper, we demonstrate that representations of the face used during the verification process should be driven by their sensitivity to these mismatches. Two representation categories of the face are proposed, parts and reflectance, each motivated by their own properties of invariance and sensitivity to different types of mismatches (i.e. spatial and spectral). We additionally demonstrate that the employment of the sum rule gives approximately equivalent performance to more exotic combination strategies based on support vector machine (SVM) classifiers, without the need for training on a tuning set. Improved performance is demonstrated, with a reduction in false reject rate of over 30% when compared to the single representation algorithm. Experiments were conducted on a subset of the challenging Face Recognition Grand Challenge (FRGC) v1.0 dataset.