Statistical Analysis of a Deep Learning Based Trimodal Biometric System Using Paired Sampling T-Test

Biometrics of individuals is distinctive to them and is constant over the course of a person's lifespan. A few trimodal biometric systems involving passive biometrics have adopted Convolutional Neural Network (CNN) technique for image recognition but the technique still suffers problems of overfitting and generalization especially with fewer datasets. Hence, the aim of this paper is to address the problems utilizing improved deep learning strategies namely; CNN-MA and CNN- MMA, and to also carry out their statistical analysis utilizing paired sampling t-test. Five hundred and seventy (570) coloured ear images, and one thousand one hundred and forty (1140) grayscale face and iris images of 190 individuals were captured through the use of a Techno F1 digital camera and CMITECH DMX-10 CCD face-iris camera, respectively. One thousand and twenty-six (1026) samples were used for training while six hundred and eighty-four (684) were used for testing. The system implementation was done on MATLAB 2016a. The statistical t- test analysis between CNN-MA and CNN-MMA showed statistical significance in terms of FPR (p = 0.014), P (p = 0.013) and RA (p = 0.005). The statistical t-test analysis between CNN- MA and CNN also showed statistical significance in terms of FPR (p = 0.000), SEN (p = 0.014), P (p = 0.000) and RA (p = 0.036). CNN-MMA and CNN-MA techniques yielded better FPR, SEN, SP, P, and RA than the CNN.