Finger-print based human verification system

The diploma thesis presents an algorithm for verification based on fingerprints. In order to achieve a simple and modular design, the algorithm is divided into number of steps. As an input, the algorithm takes greyscale fingerprint images. First, segmentation is performed where the background is separated from the area which represents the fingerprint. This is followed by the calculation of orientation field of the fingerprint using the gradient method and local frequency estimation. Both values are used as an input to Gabor filters in the step of improving the quality of fingerprint images. The enhanced image is then binarized, which is followed by the thinning procedure. In the latter step, the skeleton of the fingerprint is obtained with the help of 21 rules for the thinning of the lines and 4 rules for the thinning of the diagonal lines, where every ridge is one pixel wide. In the step of minutiae extraction, ridge endings are discovered using the method of crossing number, and ridge bifurcations are extracted using 24 masks. In the classification step, singular points are obtained by using the curvature map. With the help of singular points we determine the type of the fingerprint and the location of the reference point. Finally, the reference point orientation is calculated. With respect to the corresponding reference point all minutae are converted to polar coordinate system and a final assessment of equality of two fingerprints is given. The algorithm has been tested on four test sets used at the FVC 2002 competition. Each set contained 800 fingerprint images of different qualities and sizes. Comparison between the results obtained using the described algorithm and results from the competition FVC 2002 showed that the algorithm gives satisfactory results, but needs some improvements. Most problems appear at images of poor quality, where it is difficult to determine the course of the ridges. We successfully connected the application with the fingerprint reader. The results indicated that the application is not yet suitable for the environments where 100% reliability is necessary, further improvements are needed in the processing time as well