Modeling the Uncertainty in Finger-Vein Authentication by the Gaussian Mixture Model

The robustness and uniqueness of finger-vein makes it an ideal biometric feature for personal authentication. General finger-vein authentication methods consist of two parts, feature extraction and feature matching. Finger-vein images captured by infrared device are subject to uncertainties caused by various temperature, irregular illumination and finger posture deformation. Uncertainties cause severe artifacts, which make the extracted features unsatisfying and hard to match. We try to alleviate the problem during matching by modeling the extracted features as Gaussian Mixture Model (GMM). In the proposed method, given two feature maps of finger-vein, we first model inputs as GMM using the normal distribution transform, and then minimize the distance between two GMMs based on gradient descent, lastly we output the possibility that two feature maps belong to one person. To show its superiority, we replace conventional feature matching schemes with proposed method and test the performance gain based on two kinds of finger-vein features: finger-vein trajectory and finger-vein skeleton. Experimental results on the RATE dataset show that the proposed method is superior to the conventional methods in precision.