Deep discriminative representation for generic palmprint recognition

Abstract

State-of-the-art palmprint recognition methods have achieved significant performances. However, most of the existing methods are focused on particular scenarios such as a specific illumination or being captured using a contact-based or contactless device. Therefore, these algorithms cannot meet the ever-changing complex application requirements. To resolve this issue, this paper proposes a generic framework to represent high-level discriminative features for multiple scenarios in palmprint recognition with learned discriminative deep convolutional networks named deep discriminative representation (DDR). We propose to learn discriminative deep convolutional networks with limited palmprint training data, which is utilized to extract deep discriminative features. Then, the collaborative representation based classifier is implemented for palmprint recognition, which is flexible and practical in numerous scenarios. The experimental results demonstrate that DDR produces the best recognition performance in generic palmprint recognition compared to other state-of-the-art methods. For contact-based palmprint recognition under different lighting sources, DDR achieved the best performance on the PolyU Multi-spectral database with M_R, M_B, M_G and M_NIR, respectively. As for contactless palmprint recognition, DDR obtained the highest results on the IITD and CASIA databases.