Application Of Affinity Propagation for Prototype Sample Detection, with Application to Face Recognition

Abstract

The use of cluster centers obtained from methods based on Nearest Neighborhood, Gaussian Mixture Model and K-means/mean-shift techniques as prototype-vectors/codebook-vectors for describing a set of high dimensional data require a robust data set or a prior knowledge of number of modes present. Hence methods based on these techniques are not reliable to describe sparse high dimensional data originating from different independent sources. This paper shows the use of a data domain description method, inspired by the recently proposed clustering approach by Frey and Dueck, called Affinity Propagation. Experimental results show that the data domain description vectors obtained in a low dimensional feature space is well suited for identification of frontal face datasets (high dimensional sparse data) when compared to traditional method of Gaussian mixture model.