A Less-constrained Sclera Recognition Method based on Stem-and-Leaf Branches Network

Abstract Recognition based on the vasculature patterns observed on the sclera is a topic with great potential in biometrics. The emergence of neural network offers opportunities for more reliable identification. However, the annotation labels, which are extremely important for network learning, are often difficult to obtain, especially for sclera vessels. This paper proposes a robust sclera recognition method consisting of a non-learning-based segmentation step and a learning-based classification step, minimizing the reliance on supervision. The novelty of the approach lies in the designed stem-and-leaf branches network, named SLBNet, aggregating different levels of global and local features to obtain discriminative descriptors. Besides, a new dataset focusing on the sclera vasculature without iris, named ScleraVO, is built with the designed acquisition system. Experimental results demonstrate that the proposed recognition method achieves state-of-the-art performance and presents a good generalization ability.