Biometric Recognition Using Phonocardiograms and Convolutional Neural Networks

Data security is of real interest in many areas. As technology is constantly evolving, methods of securing data are becoming increasingly powerful and secure. Biometric recognition methods are superior to traditional recognition methods (passwords or tokens). They are not transferable, so access will be granted to the individual who presents the required unique characteristics. This paper presents two methods for biometric person recognition based on phonocardiography and 2D convolutional neural networks. The proposed methods consist of generating spectrograms for the intervals extracted from phonocardiograms and classifying them with architectures such as MobileNet, EfficientNetB0 and DenseNet121. Twenty-seven phonocardiograms were used, which are available in the CirCor DigiScope Phonocardiogram Dataset on the Physionet platform. Based on the analyses of the performance parameters, both proposed methods prove to be effective in biometric recognition of persons. For both of them an accuracy of more than 94
^%
is obtained. The method based on S1S2 interval isolation proves to be the best performing. For this method the highest accuracy of 98% was obtained for the DenseNet121 model.