Face recognition using average example image

Abstract

A novel face recognition approach by the use of a bundle of example images is presented, in which a combination of useful examples for face recognition is selected, and multiple augmented example images are made using those images. For this purpose, first, example images are divided into multiple groups using unsupervised methods such as the K-means method, and then clustering them into several groups according to their image variations. In each group, the most similar example images to an input image are found and they independently make an example average image from both the retrieved examples and their rank orders. When comparing similarities between two images, now local distance information can be utilised, which are calculated via corresponding average example images, as well as global distance information, the traditional direct comparison. The proposed approach is compared with the traditional direct comparison method on a multi-PIE database, and the generalisation ability of the proposed method across different features, such as local binary pattern, histogram of oriented gradients and Gabor features, is demonstrated.