Template Selection and Update for Biometric Recognition Systems with Nearest Neighbor Classifier

Template selection is a fundamental problem in biometric recognition systems using prototype and nearest neighbor classifier. The problem is challenging because such a small number of templates must be representative of a large amount of intra-class variations such as posture changes and lighting conditions. In addition, the templates are expected to be adaptive to small gradual or sudden changes like aging and scars. To this end, existing techniques propose to perform template update by periodically selecting new representative templates. However, they are either ineffective because selected templates are less representative, or prone to selection errors due to the presence of outliers. In this paper, we propose a novel template update framework and a robust template selection method. We use sample pass table and template pass table to accumulate information for outlier removal and representative templates selection, and further propose a novel criterion to select the optimal template set according to current templates and similarity threshold. In order to confirm the effectiveness of our method, we have carried out a number of experiments on synthetic and real datasets. The results on synthetic dataset show that our template update framework is more robust than existing ones, and at the same time. achieves a higher accuracy according to the specified performance indicator. The results on real palmprint dataset also demonstrate the superiority of our method to random selection.