A Fingerprint Liveness Detection Algorithm Based On Deep Convolution Neural Networks

As a typical biological identity authentication technology,fingerprint recognition has been widely used in many fields.However,with the continuous innovation of fake fingerprint materials and technologies,fingerprint biometric detection technology is also facing more severe challenges.The problems of fingerprint identification system security and personal privacy protection have attracted widespread attention.The traditional fingerprint liveness detection algorithm mainly relies on prior knowledge to artificially extract and classify the shallow features of fingerprints.But the extracted features are lacking the ability to represent information because of the small number and single type,which results in the poor generalization ability of the traditional fingerprint liveness detection algorithm.The method of fingerprint liveness detection based on the convolutional neural network can automatically extract the abstract deep features of the fingerprint and effectively solve the cumbersome and inefficient problems of the traditional methods.However,common convolutional neural networks also have some deficiencies.Researching and solving these problems has very important significance for the fingerprint liveness detection.In this paper,we mainly construct a fingerprint living detection framework based on deep convolutional neural network,and deeply study the difficulties faced by deep structure convolutional neural networks,including the gradient dispersion and the over-fitting caused by lack of training samples during the training process,as well as the effect of different size input samples on network validity.In order to solve the above problems,this paper mainly improves the performance of fingerprint biometric detection algorithm through the following improvements:1.In view of the existing the gradient dispersion and over-fitting phenomenon in convolutional neural network,a new fingerprint detection algorithm called F-net is proposed.The algorithm uses BN layer,inception level and global mean pool level to optimize the network,so as to reduce the number of parameters in F-net network and computational complexity.The experimental results show that the proposed F-net algorithm has high recognition rate and real-time performance.2.Aiming at the loss of important information brought by clipping or wrapping training samples,a multi-layer cascaded two-scale pooling model called MCTP-net is proposed.Two-scale pooling used in the algorithm can obtain different scale information of image features,as well as can get the output images of the same dimension for input images with different sizes,so that the network can only guarantee the validity of an image of one input size.We proposed the multi-layer cascade structure using shortcut connection technology can further obtain image features of different scales and different levels,enhance the reuse of features,minimize the loss of image information due to the conventional cascaded pooling operation,and widen the image feature expression.