Fingerprint Classification using Deep learning

Abstract

Biometrics are being extensively used in person authentication; while spoofing is used by the imposters to crack into the biometric systems. The paper deals with the fingerprint image classification into two classes viz. fake or genuine using Convolutional Neural Network (CNN) on ATVS-FFP dataset of fingerprint images of 17 users. The dataset is divided into two parts named as DS_WithCooperation and DS_WithoutCooperation, both the parts contain fake and original fingerprint images. These differ with respect to the acquisition of fake fingerprint which was done with and without the consent of the users. Thus, the fake fingerprint of latter part of the dataset were of low quality and represent the real-world scenario. The images were segmented to get the ridge region from the background noise using morphological image processing methods. The segmented images were then randomly rotated at different angles and were resized to 170X170X1. In this way the DS_WithCooperation resulted into 3264 images from a total of 816 fake and real images, similarly DS_WithoutCooperation resulted into 3072 images from a total of 768 fake and real images. This data set was then split in 3 to 1 ratio to form train and test datasets. For preparing the proposed model to classify fake and genuine images, CNN was trained using the Train data set. The model gave ACE (Average Classification Error) ranging from 0 to 2.45 on test datasets of both types with and without cooperation which is comparable to the state-of-the-art.