Cross-dataset Image Matching Network for Heterogeneous Palmprint Recognition

Palmprint recognition is one of the promising biometric technologies. Many palmprint recognition methods have excellent performance in recognition within a single dataset. However, heterogeneous palmprint recognition, i.e., mutual recognition between different datasets, has rarely been studied, which is also an important issue. In this paper, a cross-dataset image matching network (CDMNet) is proposed for heterogeneous palmprint recognition. Feature representations specific to a certain domain are learned in the shallow layer of the network, and feature styles are continuously aligned to narrow the gap between domains. Invariant feature representations in different domains are learned in the deeper layers of network. Further, a graph-based global reasoning module is used as a connection between the shallow and deeper networks to capture information between distant regions in palmprint images. Finally, we conduct sufficient experiments on constrained and unconstrained palmprint databases, which demonstrates the effectiveness of our method.