Fingerprint Match Based On Hidden Markov Models

Fingerprint identification is one of the most widely used biometric technology. As the core of fingerprint identification, many different fingerprint matching algrithms have been proposed. The algorithm based on Hidden Markov Model has attracted the most attention in recent years, but there are still some deficiencies waiting for solution. For overcoming some deficiencies in existing algorithm, a thorough study on several key points of the existing algorithm has been done and several improvements have been provided in this paper. The detail is as follow.Firstly, as for fingerprint feature selection, the existing algorithm doesn't make the best of the fingerprint feature information. For overcoming this defect, a new feature vector construction method was provided in this paper. The new vector was composed of orientation information, frequency information and curvature information. Since the new vector contained more information, it had more powerful description ability on fingerprint texture changes and laid a foundation for improving the recognition correct rate.Secondly, a new core point alignment algorithm based on identification mask was provided in this paper. The whole process was divided into two stages. The first stage was to find the alternative points. The second stage was to make sure the real core point using the identification mask. In addition, the position of the feature window was adjusted grounding on the distribution of feature information. As the result, there were more feature information in the feature window. It was helpful to enhance the accuracy of the matching algorithm.Thirdly, a new model was provided at the part of model construction, which was composed of five one-dimensional hidden Markova models. This model had more powerful ability on feature description than the existing one. Finally, the provided algorithm was realized with Visual C++ and Matlab 7.0. And the algorithm was verified with the FVC2004 fingerprint data base. Tests and analyzing showed that the provided algorithm had higher recognition correct rate than the existing one.