Local gaussian difference extrema pattern: a new feature extractor for face recognition

Abstract

In this paper, a contemporary method for face recognition i.e. local gaussian difference extrema pattern (LGDEP) is proposed. The proposed method collects the most prominent directional edges present in the image, which in turn extracts the most prominent edge information. The basic local binary pattern (LBP) encodes the central pixel with reference to its surrounding pixels in an image while the directional local extrema pattern (DLEP) encodes the directional edge information featured by local extrema in 0°,45°,90° and 135° in an image. Whereas the proposed method encodes the most prominent directional edge information by integrating the local extrema concept with the multi-resolution (three resolutions) Gaussian filter banks. The Gaussian filter banks make the method more robust as it extracts only the most prominent edges and ignores (smoothens) the minor variations due to noise. Finally, k-nearest neighbors (KNN) classifier is used for the classification. The performance of the proposed method is evaluated on three different sets of FERET Database and compared with the standard block based LBP (BLBP). After thorough analysis, it has been found that the proposed method significantly improves the results from 95% (BLBP) to 100% in terms of recognition rate.