Completely Contactless Finger-Knuckle Recognition using Gabor Initialized Siamese Network

摘要

This document presents a novel approach for contactless finger-knuckle biometric modality using Gabor initialized Deep Siamese Network. For feature extraction of finger-knuckle creases, Gabor filter is used in convolutional layer of twin CNN of Siamese Network. Validation of the model uses N-way One S hot Learning technique. A database of 146 different subjects was recorded using a smartphone camera. It contains 5 different dorsal finger-knuckle images of right-hand index finger of each individual. Experimental results show an accuracy of 94.6% and a fast convergence rate of model, which illustrate the ease of use of finger-knuckle biometrics in online applications, specifically involving smartphones, laptops and other real-time systems involving biometric verifications.