Research On Several Ensemble Learning Methods And Their Applications In Fingerprint Recognition

Until now, fingerprint recognition is still a hot research topic. Fingerprint recognition has been used in many fields, but there are many existing problems, such as high security application problem, multi-template ensemble problem, full matching score sequence problem, etc. For above problems, this thesis proposes some proper methods to deal with them. In this article, we focus on the following three topics:(1) High security application problem. The high security application, such as the access control of ATM, nuclear power stations and military secrets, etc, they all want to reach an extremely low false accept rate and as low as possible false reject rate at the same time, which is called Double Low problem. Generally, we always prefer to get a low equal error rate (EER), it can’t solve the double low problem. Because for an automatic fingerprint identification system (AFIS), it maybe has an extremely lower EER, but the false reject rate may be high when the false accept rate is extremely low. For dealing with the problem, using the ensemble learning, this thesis proposes a hybrid ensemble method. The method uses two basic fingerprint identification methods which have complementary:minutiae-base fingerprint identification and ridge-based fingerprint identification, high thresholds for two methods are set, and the false accept rate is zero. First the serial ensemble method is used, and then the parallel ensemble method is used. The effectiveness of our method is validated on FVC2002DB1and FVC2002DB2standard fingerprint database.(2) Multi-template ensemble problem. In enrollment stage, we always obtain many fingerprint images. There are two problems:first, the storage space is wasted; second, the recognition time is wasted. So, there are two challenges:(1) how to choose the proper templates for ensemble;(2) how to use the multiple templates information effectively. For dealing with the challenges, in this article, we propose a framework of multi-template ensemble for fingerprint verification. First, in enrollment stage, we propose a new template selection method, which could select template fingerprints that better represent a finger. So we can use a little of images to replace all enrollment images. Then we propose a matching score mapping to polyhedron method, which could use the similarity score among multiple templates and the similarity score between a query image and the templates effectively. The effectiveness of our method is validated on FVC2004standard fingerprint database.(3) Full matching score sequence problem. In identification stage, a query (inputted biometric trait) from a user will be compared with the templates of all users in the database, and then the system will output a matching score sequence. In existing methods, usually only the maximum matching score of the sequence is selected to determine the identifier of the query. Other matching scores of the sequence are regarded as useless. But this is not necessarily true. In practice, in addition to the maximum matching score, other matching scores of the sequence can also contribute to identification results. Based on this idea, in this paper, we define a new feature:full matching score sequence. And then, an identification method is proposed using the new feature. The effectiveness of our method is validated on our construction special fingerprint and face databases.Above methods, we only do a preliminary experiment result and don’t have a theoretical proof. So, the future work will focus on two aspects:first, using a theory to prove it; second, extending the experiments and proving the generalization of proposed methods.