Joint discriminative feature learning for multimodal finger recognition

Abstract

Recently, finger-based multimodal biometrics, due to its high security and stability, has received considerable attention compared with unimodal biometrics. However, existing multimodal finger feature extraction approaches separately extract the features of different modalities, at the same time ignoring correlations among these different modalities. Furthermore, most of the conventional finger feature representation approaches are hand-crafted by design, which require strong prior knowledge. It is therefore very important to explore and develop a suitable feature representation and fusion strategy for multimodal biometrics recognition. In this paper, we proposed a joint discriminative feature learning (JDFL) framework for multimodal finger recognition by combining finger vein (FV) and finger knuckle print (FKP) patterns. For the FV and FKP images, we first established the informative dominant direction vector by convoluting a bank of Gabor filters and the original finger image. Then, we developed a simple yet effective feature learning algorithm, which simultaneously maximized the distance of between-class samples and minimized the distance of within-class samples, as well as maximized the correlation among inter-modality samples of the within-class. Finally, we integrated the block-wise histograms of the learned feature maps together for multimodal finger fusion recognition. Experimental results demonstrated that the proposed approach has a better recognition performance than state-of-the-art finger recognition methods.