Fouling And Damaged Fingerprint Recognition Based On Deep Learning

With the development of information technology and the increasing demands for information security,the industry is urgently in need of more reliable identification techniques for identity authentication.Therefore,the biometric identification methods have become a compelling issue.Among the methods,the fingerprint identification technique attracts much interest when considering its excellent feasibility and reliability performance.The traditional fingerprint identification method relies on matching feature points to compare the similarities between finger prints with a very high recognition rate.However,the realization of excellent performance relys on three indispensable facts: the high quality of the fingerprint samples,the perfect fingerprint preprocessing algorithms and the accurate feature points positioning algorithms.As the quality of fingerprint samples differs from each other in which blurred,deformed or even serious defected fingerprints occasionally appear,the traditional fingerprint recognition algorithm for classification is unable to maintain its advantages.Aiming at these problems,we propose four new algorithms based on Deep Learning for damaged fingerprint recognition: FFPF_CNN(Fuzzy Feature Points of Fingerprint Identification Based on Convolutioanal Neural Network),CBF_CNN(Core Block of Fingerprint Identification Based on Convolutioanal Neural Network),damaged fingerprint identification algorithm based on merged multi block graphs and an optimized damaged fingerprint identification algorithm on account of multi feature graphs.The fingerprint recognition process can be conducted accurately without the requirements for complete fingerprints or the entire feature points of the fingerprint which proves to be a promising solution.The FFPF_CNN algorithm uses the Poincare formula to find all the existed feature points of the blurred fingerprints,and then dims these feature points by using fuzzy algorithms;the fuzzy graphs of feature points are fed to the Convolutional Neural Network(CNN)for training and recognition.This method focuses on the quantity,position,and the relationship between the feature points rather than the line information.Comparing with the traditional feature point matching algorithm,this algorithm does not require a specific amount of the feature points,but extracts features only by using a part of them.The CBF_CNN algorithm also uses the Poincare formula to find the core points of the fingerprints,and extracts a little piece of the image around the core point into the CNN model.The core block of fingerprint not only cuts a large region of the damaged fingerprints,but retains the most important area with the richest information.This method can save much manpower and material resources;it also greatly improves the recognition rate of the damaged fingerprint.We propose an improved damaged fingerprint identification algorithm by introducing the multiblocks of fingerprint on the basis of the CBF_CNN algorithm.As previously described,we just utilize the core block of fingerprint with much area aborted.This new algorithm inputs the core region and the surrounding eight areas of the fingerprint images and original fingerprint images into CNN model.In this way we can increase the recognition rate as much as possible by using most of the fingerprint feature information.Inspired by the former two algorithms,the thesis demonstrated another finger print identifying algorithm which integrated two feature images together and then inputed the integration.The detail was that the core part of the fingerprint and the fuzzy image of the feature points were stiched from the top to the button,and the stiched single one image was input into the convolution neutral networks.This algorithms is able to take full use of the information of the fingerprint images such as the feature points,strings and patterns.The experimental results show that the two proposed algorithms succeed in increasing the recognition rate of the damaged fingerprint.