| In this thesis, the author's objective is to design a robust automated fingerprint classification system(AFCS) with a higher accuracy to overcome the shortcoming of the traditional method. This AFCS includes five modules: fingerprint collection, directional image computation, image segmentation, feature extraction, and fingerprint classification. All of these have been explored deeply. Moreover, some new concepts and algorithms have been proposed as follows: (1) A novel concept on the continuously distributed directional image/field (CDDF), and the method to compute it in fingerprint images are developed. The CDDF transits smoothly, and exhibits not only good continuity, well gradualness and excellent robustness to noises, but also very high precision. So it can represent the basic structural feature of fingerprint more precisely. Therefore, it can not only solve the key problem of continuity, accuracy and robustness of fingerprint directional fields, but also improve the performance of AFIS effectively. (2) The classical formula to compute the Poincaré Index is improved so that the modified version of Poincaré Index can present not only the rotation angles, but also the rotation direction of the vector in the vector field, exactly. (3) A new algorithm for fingerprint singularity detection based on both the CDDF and the modified Poincaré Index is proposed, which can locate the singularities (core points, and delta points)at pixel level with an accuracy of only one pixel. (4) A novel fingerprint classification algorithm based on both the CDDF and the modified Poincaré Index is developed finally, which is invariant to image rotation of any angles, and successfully solves the problem of image rotation, translation, and transformation in fingerprint classification. A classification accuracy of 97.05% has been achieved. So it has better classification performance than that previously reported in the literature. |
PDF Full Text Request