EEPNet: An efficient and effective convolutional neural network for palmprint recognition

Abstract

Palmprint recognition is an important biometrics technology. In recent years, a lot of palmprint recognition methods based on convolution neural networks (CNNs) have been proposed. However, existing CNNs specially designed for palmprint recognition have high computational complexity. In order to make palmprint recognition method based on deep learning work well on mobile devices, lightweight neural networks must be used. However, up to now, there is very little research on this topic. In this paper, we propose an efficient and effective palmprint recognition network (EEPNet), which is a lightweight neural network. EEPNet is designed based on MobileNet-V3, and further compresses the number of layers and enlarges the convolution kernel. In addition, we design two new loss functions including Balanced Loss and Contrast Loss. Balanced Loss is suitable for various specific data sets, while Contrast Loss can achieve the purpose of training difficult samples without manual parameter adjustment. According to the characteristics of palmprint recognition, we add five strategies to improve the recognition performance including image splicing, image dimension reduction, data augmentation, cascade channel attention mechanism, and hard case mining mechanism. We conduct thorough experiments on seven palmprint databases. The experimental results show that the overall recognition performance of our method outperforms classic and state-of-the-art palmprint recognition methods on the palmprint databases with normal quality. We compare our method with other CNNs in four aspects: precision, speed, parameter quantity and FLOPs. The experimental results show that our method is more efficient and has high recognition accuracy.