Robust multimodal multivariate ear recognition using kernel based simultaneous sparse representation

摘要

In this paper, we propose a novel multivariate multimodal ear recognition method which exploits correlation between left and right ear modality of an individual for his/her identification using joint sparse representation and its variant, joint dynamic sparse representation based classification approach. To make the problem much more robust against outliers that might be resulted from illumination variation or noises due to inaccurate measurements or from partial occlusion due to hair or ornaments — especially for female subjects, we employ a novel weighted multivariate regression scheme under joint sparse as well as joint dynamic sparse penalization. That particular scheme learns a set of weights iteratively for each and every residual corresponding to each observation and subsequently, during the time of classification, gives lesser weight to elements detected as outliers such that they are not able to participate for query set representation. To further improve accuracy of the system, the proposed method is kernelized to tackle non-linearity infusion made by pose variations and occlusions. In the end, extensive experimentations are carried out over a novel database developed in our laboratory to compare performance of the proposed method to several competitive, state-of-the-art methods in order to check suitability of the proposed classification method for various real life applications.