Boosting Unconstrained Palmprint Recognition with Adversarial Metric Learning

With great potential and prospects in the market, palmprint recognition has also aroused wide concern in the academic community. However, there are yet some burning issues needed to be explored. On the one hand, most public palmprint databases are collected with exquisite equipments in a semi-closed environment. But it is very inconvenient for users in practical application. On the other side, most deep metric learning methods in literatures ignore the distribution of palmprints in the global feature space, which may result in severe locality and unbalance. In order to address the above problems, we novelly established a large-scale unconstrained palmprint database called XJTU-UP, which consists of more than 20,000 images collected by 5 different mobile phones. Most importantly, we proposed an adversarial metric learning methodology to make different categories of palmprints uniformly and dispersedly distributed in the hypersphere embedding space. Vividly speaking, the joint metric term and confusion term are like 'extensional and compressive deformation of the helical springs', which encourages to learn equidistributed representations. Finally, the experimental outcomes on 8 representative palmprint databases unfold the superiority of our algorithms-the palmprint identification accuracy has a maximum increase of 4% and the verification performance is greatly boosted by up to 15%.