A General Weighted Multi-scale Method for Improving LBP for Face Recognition

Abstract

LBP (Local Binary Pattern) is a popular image descriptor (feature) that has been widely used in face recognition. LBP has some parameters, and different parameter values leads to different LBP feature vectors. In practice usually only one feature vector is used for one image, thus information about image content is not utilised fully by LBP. In this paper a novel way of utilising LBP features more fully is presented. Different LBP feature vectors are extracted for one image, corresponding to different combinations of LBP parameter values. These vectors are weighted and used in a distance function. Then the k-nearest neighbour classifier is used. Experiments have been conducted on the AR database. Results show this method does indeed produce better classification performance, suggesting that more information considered this way can have values.