Fusion of speech and face by enhanced modular neural network

Abstract

Biometric Identification is a very old field where we try to identify people by their biometric identities. The field shifted to bi-modal systems where more than one modality was used for the identification purposes. The bi-modal systems face problem related to high dimensionality that may many times result in problems. The individual modules already have large dimensionality. Their fusion adds up the dimensionality resulting in still larger dimensionality. In this paper we solve these problems by the introduction of modularity at these attributes. Here we divide various attributes among various modules of the modular neural network. This limits their dimensionality without much loss in information. The integrator collects the probabilities of the occurrences of the various classes as outputs from these neural networks. The integrator averages these probabilities from the various modules to get the final probability of the occurrence of each class. This averaging is performed on the basis of the efficiencies of the modules at the time of training. A module that is well trained is hence expected to give a better performance than the one which is not well trained. In this manner the final probability vector may be calculated. Then the integrator selects the class that has the highest probability of occurrence. This class is returned as the output class. We tested this algorithm over the fusion of face and speech. The algorithm gave good recognition of 97.5%. This shows the efficiency of the algorithm. © 2010 Springer-Verlag Berlin Heidelberg.