Applications of Different CNN Architectures for Palm Vein Identification

Abstract

In this paper a palm vein identification system is presented, which exploits the strength of convolutional neural network (CNN) architectures. We built and compared six different CNN approaches for biometric identification based on palm images. Four of them were developed by applying transfer learning and fine-tuning techniques to relevant deep learning architectures in the literature (AlexNet, VGG-16, ResNet-50 and SqueezeNet). We proposed and analysed two novel CNN architectures as well. We experimentally compared the identification accuracy and training convergence of these models. Each model was trained and evaluated using the PUT palm vein near infrared image database. To increase the accuracy obtained, we investigated the influence of some image quality enhancement methods, such as contrast adjustment and normalization, Gaussian smoothing, contrast limited adaptive histogram equalization, and Hessian matrix based coarse vein segmentation. Results show high recognition accuracy for almost every such CNN-based approach.