Face verification using elastic graph matching based on morphological signal decomposition

Abstract

In this paper a novel method for frontal face verification is proposed. It is based on the morphological signal decomposition, a procedure that is used to model a facial image region as a sum of components. During the procedure, partial sums of components create a sequence of reconstructed images which starts with a crude approximation of the facial region that recursively becomes finer. More specifically, a feature vector is created at each node of a sparse grid superimposed on the facial area by concatenating the gray level values of the reconstructed images at the grid node position. When a candidate person claims the identity of a reference person, a variant of dynamic link matching, the so-called morphological signal decomposition—dynamic link architecture, is applied to yield a matching error between the reference grid and a variable grid that is built over the facial image region of the candidate person. Local coefficients are derived to weigh the contribution of each node to the total matching error according to the node discriminatory power. Moreover, an analysis of the discriminatory power of each level in morphological signal decomposition is undertaken to assess better the behavior of the proposed method. Experimental results are reported on the M2VTS facial image database yielding a very low equal error rate.