Fingerprint identification technology was first adopted by law enforcement agencies and tremendous success has been achieved due to stability and exclusion of fingerprint. Growing identity fraud, as well as widely acceptance by the public has all ushered in an era of fingerprint-based person identification applications in civilian domains. In this thesis, we analyze some problems in our embedded system and investigate some of key issues in fingerprint image reconstruction, enhancement and matching. The main contributions are as follows: 1. A new algorithm for fingerprint reconstruction based on motion estimation is proposed. We select reference block dynamically and feedback prediction of motion vector using motion estimation theory and physical law of finger scanning over the sensor. Multi-frame and sub-pixel motion estimation are integrated to obtain relative displacement of successive slices to reconstruct fingerprint image. Experiments have shown that our algorithm can work well in real-time systems. 2. A new binarization method based on geometric properties of enhanced fingerprint is proposed. We propose an approach to fingerprint binarization based on the trace of Hessian and verify that it is equivalent to the method based on maximum principle curvature on the assumption of orientation coherence in local structure. Therefore, the anisotropic diffusion filter is used to adapt to the coherence structure of fingerprint. Experimental results show that the algorithm performs better than traditional algorithm. 3. A new algorithm for quantization and compression of orientation field and mutual information based matching is presented. The algorithm of run length coding is improved and applied to fingerprint template compression according to the relativity of orientation field and technical characteristics of DSP systems. We match two fingerprint images using mutual information theory by two random variables of template orientation map and query fingerprint orientation map. The approach to fingerprint matching can obtain the balance for recognition performance and compression efficiency. 4. Two matching algorithms are proposed combining minutiae triangular feature with orientation feature on the basis of data level and decision level respectively. Firstly, triangular feature vectors immune to rotation and transformation are extracted and consulted for matching without alignment. To decrease time consumption, we propose two approaches to searching for equivalent triangles in distributed areas and geometric transformation parameter clustering. Secondly, we propose a new matching framework cascading minutiae triangular matcher and orientation matcher., so the better performance can be obtained when the second algorithm is used and combined with the first algorithm with a small probability. The algorithms presented in this thesis have been applied to our recognition systems and good test results have been obtained.
修改评论