Distance metric learning for recognizing low-resolution iris images

Low-resolution (LR) iris images are inevitable, especially in the iris recognition systems under less constrained imaging conditions which are desirable to extend the applicability of iris biometrics. It is a challenging problem to match LR probe iris images with high-resolution (HR) ones captured at enrollment stage. This paper presents a heterogeneous metric learning algorithm which can favorably improve the accuracy of LR iris recognition. The basic idea of the method is to learn an appropriate distance metric to transform the heterogenous (LR vs. HR) iris matching results towards the desirable homogeneous (HR vs. HR) ones and then further enhance the separability between intra-class and inter-class matching samples. This learning procedure not only utilizes label and local information, but also fully exploits the sample correspondence and the ideal application scenario as the specific prior information. Two steps are included in the proposed method. Firstly, the ideal pairwise similarities are defined on the training set to faithfully achieve the basic idea above. Secondly, the Mahalanobis distance is learnt by minimizing the divergence between the matching results measured by the target Mahalanobis distance and the ideally defined matching results. Extensive experiments show that the proposed metric learning solution consistently outperforms state-of-the-art metric learning methods and can further enhance the performance of existing LR iris recognition approaches.