An Approach to Pulse Coupled Neural Network Based Vein Recognition

Abstract

Hand vein recognition has received increasing attention in biometric identification for the uniqueness, stability, and easiness of collection of the vein image. Local Binary Pattern (LBP) has been a widely used texture descriptor, being well developed in vein recognition. However, the use of histogram in LBP as the feature of the vein image leads to the loss of global spatial information. That loss results in final accuracy reduction for absence of geometry structure, being essential to vein representation. In this paper we use Pulse Coupled Neural Network (PCNN) to process original LBP feature map as a solution to spatial information loss, because in PCNN pulse emitting and spreading reflects the intensity distribution pattern related to the vein geometry. The image time signature (image icon) produced by pulse emitting and spreading in PCNN is used as the vein feature. Using PCNN to extract time signature, we adopt a multi-valued linking channel in each neuron of the network to control the neighboring influence more precisely, and we introduce an adaptive linking strength ($\beta$) to address the less information on vein pattern of the dark region in the vein image. As for the vein binary representation, Unit-linking PCNN is employed. It uses pulse spreading to perform eroding operation on the K-clustering binarization result of the vein image. After computing the time signature and the binary representation, Support Vector Machine (SVM) fusion strategy is used to fuse them. The EER (Equal Error Rate) of our approach is 0.03% on the CASIA Multi-spectral Palmprint Image Database, which is better than four other approaches, including the multi-sampling method, mutual foreground LBP (MF_LBP) method, and so on.