Research On The Fingerprint Liveness Detection Based On Extreme Learning Machine

With the wide application of fingerprint recognition technology,more and more criminals are using fake fingerprints to attack recognition systems,access to legitimate users' rights illegally,that has posed a serious threat to the privacy information and property of legitimate uses.In order to solve the above problem effectively,we focus on the problem of fingerprint liveness detection in this thesis.Furthermore,we place emphasis on the software based fingerprint liveness detection method to reduce hardware cost and increase deployment flexibility.Method based on software uses image processing algorithm to extract the liveness features directly from the collected fingerprint images for the identification of real and fake fingerprints.This method eliminates the need to add additional hardware and simply updates the software architecture of the original system.In this thesis,we treat the fingerprint liveness detection problem as a two-category classification problem,so we study from both the classifier selection and the feature extraction techniques.For the problem of classifier selection,the thesis breaks through the limitations of the traditional support vector machine(SVM)algorithm in the field of fingerprint liveness detection,and uses a new classification learning algorithm called extreme learning machine(ELM),which is simpler and has better generalization performance.For the problem of feature extraction,we proposed two improved feature extraction algorithms based on the previous studies.One is the improved third-order co-occurrence matrix based on gradient(Improved-TCMG),and the other is the third-order co-occurrence matrix based on Laplace of Gaussian(TCM_LoG).The first algorithm has two improvements.(1)In the process of feature extraction,the ridge region segmentation is added,which can remove the background of the fingerprint image,reduce computational complexity and avoid the interference of the irrelevant features in the background region.(2)In gradient calculation,we use Sobel operator instead of the difference operator,which not only can smooth the noise contained in fingerprint image to a certain extent,but also make the calculated gradient value more accurate.The second algorithm is inspired by the first algorithm.In the feature extraction process,the co-occurrence matrix is calculated based on the second derivative of LoG instead of the gradient.This has two advantages.The LoG operator can perform Gaussian smoothing on the fingerprint image and suppress the noise added during the collection.In addition,it can make the extracted liveness features more accurate.The LoG operator has no directionality compared to the gradient operator,and only one co-occurrence matrix is calculated,which can greatly reduce the computational complexity.The paper carried out comparative experiments on the LivDet 2011 fingerprint database.The experimental results are as follows.(1)With respect to the problem of fingerprint liveness detection,ELM has fewer adjustment parameters,shorter training time,and better generalization performance than SVM.(2)The Improved-TCMG algorithm can achieve a lower average classification error rate than the current popular algorithm.(3)Compared with Improved-TCMG,TCM LoG has slightly lower performance,but its feature extraction speed is 2 times higher,that could improve computing efficiency significantly.