Difference-Based Local Gradient Patterns for Image Representation

Abstract

This paper aims to examine the impact of pixel differences on local gradient patterns (LGP) for representing facial images. Two difference-based descriptors are proposed, namely, the angular difference LGP (AD-LGP) and the radial difference LGP (RD-LGP) descriptors. For evaluation purpose, two experiments are conducted. The first is face/non face classification using samples from CMU-PIE and CBCL databases. The second is face identification under illumination variations using the extended Yale face database B and the CMU-PIE face database. The experimental results show that both descriptors demonstrate, generally, a higher capability in discriminating face patterns from non-face patterns than the standard LGP. However, in face identification, the AD-LGP descriptor shows robustness against illumination variations, while the performance of the RD-LGP descriptor degrades with hard illuminations. Furthermore, we enhance the RD-LGP descriptor using the Average-Before-Quantization (ABQ) approach in order to increase its robustness toward illumination changes.