CNN Approaches for Dorsal Hand Vein Based Identification

Abstract

In this paper we present a dorsal hand vein recognition method based on convolutional neural networks (CNN). We implemented and compared two CNNs trained from end-to-end to the most important state-of-the-art deep learning architectures (AlexNet, VGG, ResNet and SqueezeNet). We applied the transfer learning and fine-tuning techniques for the purpose of dorsal hand vein-based identification. The experiments carried out studied the accuracy and training behaviour of these network architectures. The system was trained and evaluated on the best-known database in this field, the NCUT, which contains low resolution, low contrast images. Therefore, different pre-processing steps were required, leading us to investigate the influence of a series of image quality enhancement methods such as Gaussian smoothing, inhomogeneity correction, contrast limited adaptive histogram equalization, ordinal image encoding, and coarse vein segmentation based on geometrical considerations. The results show high recognition accuracy for almost every such CNN-based setup.