Boosting hand vein recognition performance with the fusion of different color spaces in deep learning architectures

Human recognition and authentication through biometrics generally rely on feature extraction from images of physiological traits. These images can be represented in different color models. This study represents human palm vein images in five different color spaces (RGB, XYZ, LAB, YUV, and HSV) and combines their decisions in CNN models for person recognition. Color spaces are generally represented in three channels. This study identifies the channel with the highest contribution to pattern recognition in images and proposes to use only this channel per color space in the identification process instead of all three channels. The experiments confirm that channels representing how humans perceive colors are generally mostly responsible for features extracted from vein pattern biometrics. The proposed architecture is tested using modified AlexNet, VGG-19, and ResNet-50 Convolutional Neural Network (CNN) models on palm vein datasets from the FYO, PUT, and VERA databases. Experimental results showed considerable improvement in palm vein recognition compared to similar studies.