Learning Discriminant Direction Binary Palmprint Descriptor

Lunke Fei,Bob Zhang,Yong Xu,Zhenhua Guo,Jie Wen,Wei Jia

DOI: https://doi.org/10.1109/tip.2019.2903307

IF: 10.6

2019-01-01

IEEE Transactions on Image Processing

Abstract:Palmprint directions have been proved to be one of the most effective features for palmprint recognition. However, most existing direction-based palmprint descriptors are hand-craft designed and require strong prior knowledge. In this paper, we propose a discriminant direction binary code (DDBC) learning method for palmprint recognition. Specifically, for each palmprint image, we first calculate the convolutions of the direction-based templates and palmprint and form the informative convolution difference vectors by computing the convolution difference between the neighboring directions. Then, we propose a simple yet effective model to learn feature mapping functions that can project these convolution difference vectors into DDBCs. For all training samples: (1) the variance of the learned binary codes is maximized; (2) the intra-class distance of the binary codes is minimized; and (3) the inter-class distance of the binary codes is maximized. Finally, we cluster the block-wise histograms of DDBC forming the discriminant direction binary palmprint descriptor for palmprint recognition. The experimental results on four challenging contactless palmprint databases clearly demonstrate the effectiveness of the proposed method.