Design of a Lightweight Palmf-Vein Authentication System Based on Model Compression

Palm-vein authentication is a secure and highly accurate vein feature authentication technology that has recently gained a lot of attention. Convolutional neural networks (CNNs) provide relatively high performance in the field of image processing, computer vision, and have been adapted for feature learning of palm-vein images. However, they often require high computation that not only are infeasible for real-time vein verification but also a challenge to apply on mobile devices. To address this limitation, we proposed a lightweight MobileNet based deep learning (DL) architecture with depthwise separable convolution (DSC) and adopt a knowledge distillation (KD) method to learn the knowledge from the more complex CNN, which makes it small but effective. Through the depth of separable convolution, the number of model parameters is significantly decreased, while still remaining high accuracy and stable performance. Experiments demonstrated that the size of the proposed model is 100 times less than the Inception v3 model, while the performance can go beyond 98% correct identification rate (CIR) for the CASIA database.