Qun Gao
Fingerprint Verification using Cellular Neural Network
1. Auflage/1^st edition 2003, XII, 206 Seiten/pages, € 64,00. ISBN 3-89649-894-0

In this thesis, a CNN-based fingerprint verification system is realized. It consists of three main processing stages, Image Preprocessing, Feature Extraction and Feature Matching, and a system database.
In the Image-Preprocessing stage, the quality of an original gray-scale noisy fingerprint image is enhanced. As a result, a binary thinned fingerprint is obtained. In the Fingerprint Feature-Extraction stage, distinguishable real features (ridge endings and ridge bifurcations) in the thinned fingerprint as well as their feature attributes are extracted. False features are eliminated based on a distance criterion. In the subsequent Fingerprint Feature-Matching stage, a similarity comparison scheme which is tailor-made for CNNs is presented. It is able to tackle the translation distortion inherent in fingerprint images. A special system database is proposed which takes the rotation distortion into account by storing not only the feature images of the system user's fingerprint, but also their rotated versions. This greatly facilitates the similarity comparison scheme, thus speeding up the feature matching process. For the final decision, several criteria are investigated and a combination of two simple criteria with an adjustable parameter is proposed.
The performance of the whole system, i.e., the ability of the system to verify fingerprints which belong to the system user and the ability to reject fingerprints which belong to imposters who are not justified to access the system, is evaluated by using a real fingerprint test database. Due to the great variation in quality between different fingerprints even from the same user, the concept of "optimum fingerprint version" is proposed for the enrollment mode in order to improve the False Rejection Rate of the system when only one fingerprint is allowed to be stored. If memory allows to store more than one fingerprint, the concept of "best combination" of available fingerprints is presented. Moreover, we show that through adjusting the parameter in the decision criteria, the system can be used for applications requiring high security as well as for forensic applications. In practice, the system has to trade off between the False Rejection Rate and the False Acceptance Rate in order to satisfy the requirements of a specific application.
Two issues related to hardware implementation, robustness and processing speed of the system, are addressed as well. By introducing template decomposition, the robustness of high-connectivity bipolar CNNs is enhanced, thus increasing the robustness of the corresponding stage. The price to be paid is an increase in settling time. The interdependence between robustness and settling time due to template decomposition is studied. For gray-scale CNNs (for which the definition of robustness used here is meaningless, since it would be equal to zero), a new measure, perturbation tolerance, is introduced to quantify the ability of a gray-scale CNN to tolerate template parameter inaccuracies. By definition, the robustness of a CNN is a special case of its perturbation tolerance. In addition, for the Binarization operation, the maximum perturbation of template parameters allowed is derived.
Finally, with the aim of exploiting the potential of CNNs in the area of general image processing and pattern recognition, a CNN-based rotation algorithm is designed to rotate binary images. Image rotation is realized by shifting black pixels successively along their individual paths which are predetermined in a control image. The control image specifies the shifting direction and the shifting speed of each pixel.

Direkt bestellen bei / to order directly from: Hartung.Gorre@t-online.de

Reihe "Series in Signal and Information Processing" im Hartung-Gorre Verlag