This paper reports on efforts to correct descriptor errors in minutiae, and to evaluate the effect of feature accuracy, on matching performance, through Bayesian filtering and Sequential Monte Carlo approximation; it notes that results were shown to be statistically significant and led to improved matching performance as well as the reduction of average minutia position error.
Fingerprint-based recognition relies on the matching of features derived from the ridges and valleys of the friction ridge surface. When a large quantity of good-quality features are available, identification can be made with a high level of confidence. When portions of the fingerprint image are of lower quality, accuracy will suffer in the feature extraction and subsequent matching steps. This study is concerned with the correction of descriptor errors in minutiae (ridge endings and bifurcations), and with evaluation of the effect of feature accuracy on matching performance. The approach applies Bayesian filtering to refine minutia location and direction descriptors, using Sequential Monte Carlo approximation of a joint probability distribution near each minutia. The distribution approximates the location and orientation of the minutia, given measurements on local greyscale information, from which an expectation can be determined. Experimental results have been presented which demonstrate improvement of localization accuracy with respect to ground truth data when using a well-known minutia extractor. These results are shown to be statistically significant, and to lead to improved matching performance. In addition, the authors were able to reduce the average minutia position error for a set of reference minutiae by 83 percent when using the proposed refinement method. (Published Abstract Provided)