The technology of computer and internet help us to save, share and disseminate information more conveniently and more effectively, while the demand to reliable personal identification techniques increases rapidly in consequence. The traditional identification methods through password, key and ID card etc. are unsafe because they need to be remembered or taken along and are easy to be obtained or abused by others. What’s more, they cannot ensure the consistency of the digital identity and physical identity of one person, without which the security of the information could not be protected effectively. Biometrics, which can resolve these problems fundamentally, attract more and more attention in the last several decades. While fingerprint recognition is one of the most studied and widely used biometrics. Recently great improvement has been achieved in the fingerprint sensing technology and automatic recognition algorithms. But the accuracy of state-of-the-art fingerprint matching systems is still not comparable to human fingerprint experts in many situations. One of the most important reasons is the nonlinear distortion produced in acquisition process. The wide existences of distortion in fingerprint images reduce the accuracy of matching algorithm and the performance of the system obviously, especially the distortion of fingerprint images acquired from different sensors. We define the matching of fingerprints with large distortion and fingerprints from different sensors as distorted fingerprint matching. This work focus on the study and analysis about distorted fingerprint matching. The main contribution of this thesis can be concluded as follows: 1. In the matching of fingerprints with large distortion, the similarity of local features, e.g. minutiae, are easily to be affected by distortion while the accuracy of global features is limited by their fuzziness, which lowers the overall performance of the matching system. A score-level fingerprint feature fusion algorithm with prior knowledge is proposed to resolve this problem: a series of regular processes including enhancement, feature extraction and registration are executed firstly, and several scores are calculated based on the registration; then the distribution of each scores and their contribution to the final matching score are analyzed as prior knowledge, and the fusion parameters are trained through a genetic algorithm accordingly; finally, all of the scores calculated and analyzed are fuse...
修改评论