On selecting Gabor features for biometric authentication

In this paper we describe a Gabor feature selection technique that allows to develop a fast and robust Gabor feature based biometric system. Existing Gabor based methods use a huge number of Gabor features to represent the patterns, our experiments on different biometric characteristics show that using only few (~ten) Gabor features it is possible to achieve a very low Equal Error Rate. In this work, we propose a multi-matcher system where each matcher is trained using a single Gabor Filter (with a given scale and orientation) convolved with a sub-image of the whole image, and the matchers are finally combined using the “Sum Rule”. Only a low number of Gabor Filters and sub-images, selected by running the Sequential Forward Floating Selection (SFFS), are exploited in the fusion step. The system has been tested on two biometric traits: Ear Authentication and Finger Authentication. The experimental results show the effectiveness of the feature selection in terms of Equal Error Rate and Area Under The ROC curve.