Neural Network Augmentation: Generating Palm Vein for Effective Recognition

Palm vein recognition has been developed as a preferred biometric authentication technique. Deep neural networks based recognition methods has received increasing interests recently. However, in real scenarios applications, the recognition accuracy is reduced due to the illuminated inconsistency between the real-time acquisted image and the registed image. In this paper, we propose a data augmentation method using limiting contrast histogram imbalance to enhance and expand training samples. We use the ResNet50 as backbone network, and ultilize a morphological method for fast palm locating. We conduct 864000 of experiments based on the public CASIA and PolyU databases. The EER was $0.004 \%$ and the best recognition rate was increased to 96.46%.