Local Gradient Increasing Pattern (LGIP) for facial representation and gender recognition

Abstract

A robust facial representation is an essential component for gender classification. This paper introduces a new local feature, Local Gradient Increasing Pattern (LGIP), which expresses the local intensity increasing trend. A LGIP feature is to encode intensity increasing trends in 8 orientations at each pixel using signs of directional gradient responses, and overall increasing trend is assigned with a decimal label. A facial image is partitioned into overlapping regions from which LGIP histograms are obtained and concatenated into a single feature vector. Gender classification is carried out using SVM classifier based on the LGIP-based facial descriptor. We investigate the influence to recognition rates by two factors, image resolution and person-dependent/independent condition. Experiments are performed on two replicable image sets from CAS-PEAL and FERET databases, and the results show that our method achieves better performance than many other methods. © 2012 Springer-Verlag.