Automatic fingerprint identification is a challenging interdisciplinary field, which includes image processing, pattern recognition, computer technology and so on. Currently, research on fingerprint recognition has received considerable achievements, but some critical problems in this field are still needed to be resolved. The performance evaluation of automatic fingerprint identification algorithm is one of the important issues in international conferences and journals on pattern recognition. In this thesis, we have identified and explored the performance of automatic fingerprint identification algorithm based on statistical analysis. The calculation of empirical error rates and theoretical estimation of intrinsic error rates are investigated, respectively. The contribution of this dissertation is as follows: (1) We have developed a research framework to provide some insights into the strengths and limitations of the automation in matching fingerprints. An introduction to state-of-the-art in fingerprint identification technology as well as previous research done in this field is presented. (2) A subset bootstrap algorithm to measure confidence interval of equal error rate, so that the stability performance of different automatic fingerprint identification algorithms on the same data sets can be compared. We systematically study and compare this subset bootstrap technique with the conventional parametric and nonparametric (bootstrap) methods for measuring confidence intervals. Experimental results show that the subset bootstrap method gives accurate indication of the significance of the estimates such as FMR, FNMR, and EER. (3) A statistical model based on probability is proposed to estimate the uniqueness of a fingerprint template. The model is used to assess the performance limitations of popular fingerprint verification algorithm based on the pair-mate minutiae representation of fingerprints. Results are shown using FVC2002 database. These results contribute towards making fingerprint matching a science and setting the legal challenges to fingerprints. (4) We theoretically estimate the probability of a false correspondence between two fingerprints from different fingers using the pair-mate minutiae representation ot fingerprints. Unlike the previous work, which assumes that the minutiae are independent distributed, in our model, we don not make this assumption and measure the positions and orientations of pair-mate minutiae. The rates obtained by our approach are significantly lower than that of previously published research. (5) The distribution of matching scores is analysised using some statistics such as the chi-square measure, Kolmogorov-Smirnov measure, and so on. Experimental results show that the match scores are not independently identically distributed for all subjects, and suggest that a significant part of the error is due to few fingers whose data are not i.i.d.
修改评论