Research On Several Machine Learning Methods And Their Applications In Video-Based Fingerprint Verification

Biometrics has drawn great attention in recent years. Unlike conventional possession-based (e.g., passport) or knowledge-based (e.g., passwords) identity authentication schemes, biometrics are easy to carry and cannot be misplaced,guessed, forgotten or be easily forged. As the earliest and best developed biometrics, Automatic Fingerprint Recognition Technology (AFRT) has been applied successfully in criminal investigation, judicial activities, door access system and other fields.Despite the brilliant achievements in fingerprint recognition technology, the accuracy of the present Automatic Fingerprint Identification System (AFIS) is still far from the standards in some fields of high-demand, thus becoming the prime obstacle to its wider application.Many institutions are now aiming at the higher accuracy of AFIS, for example, by using videos for fingerprint verification to explore the more information inside.There are three advantages by using video for fingerprint verification. (1) More useful information will be explored by fingerprint videos for higher accuracy of fingerprint verification; (2) User Experience in collecting fingerprint videos is identical to that of collecting a single impression. (3) The fingerprint left on the physical surface could be stolen, while the dynamic information in fingerprint video is nearly impossible to be stolen.Video-based fingerprint verification was studied, and the disadvantages in valuing the matching scores of two videos using S= SO+w·(SO-SI) were analyzed. A new idea was proposed to acquire the optimized performance of fingerprint verification. A matching result with two features SI and SO was viewed as a sample. The task of verifying whether two fingerprint videos are genuine matching or impostor matching, was converted to classification task of samples with two-dimensional features (SI,SO). And machine learning algorithms(KNN,Decision tree,SVM) were adopted to classify every matching result.Experiment results showed that the accuracy of video-based fingerprint verification was significantly improved by using machine learning algorithm compared to the results by using threshold. And the current method avoided the complex process of selecting parameters and thresholds. In the experiment, the distribution of impostor samples is relatively concentrated, while some outliers in the genuine sample set embedded into the impostor sample set result in overlapping samples These overlapping samples lead to the performance degradation of classifier and badly affect the result of the classification of samples using machine learning algorithms. To overcome this problem, improve the performance of classifier and the accuracy of video-based fingerprint verification, the overlapping samples were viewed as "outliers" of fingerprint videos. New hyperplanes could be obtained after deleting these overlapping samples. By analyzing the characteristics of these "outliers" in the genuine sample set, two outlier detection methods were proposed to deal with the training sample set. One method is to delete some samples whose value of SO is less than a threshold, and the other one is to delete some samples with lower density in the genuine sample set. Experiment results showed that the method was effective.