Multi-Modal Biometric Age Prediction Through Deep Learning Framework

Abstract

The biometric prediction of an individual by using their unique characteristics is useful in many practical situations. Traditional biometrics primary focus is on person identification. However, via soft biometric processing, different general information of a user may be exploited. A multimodal approach has a lot of potential to overcome several fundamental constraints of a unimodal scheme when it comes to security, privacy, and identification. It is observed that there has been a surge in passion for widening biometric prediction capabilities to include easily quantifiable factors like the subject's age, and more recently, advanced qualities like emotional or mental states. In this regard, multimodal biometric features of a person are used to predict their corresponding age using a deep convolutional neural network. The datasets used in this study are acquired locally using a customized hardware setup, and this database includes the palm print, palm vein, dorsal vein, and wrist vein images of an individual. The features of each biometric are extracted and fused at the score level, followed by the age prediction using a tailored deep convolutional neural network. A Sequential model is built, and Adam optimization is applied, and the loss function of the model is determined by mean squared error values. The obtained error values from the model show decent error rates for different learning percentages. The experimental results obtained using the proposed method are competitive to previous studies.