Fingerprint classification

Fingerprint classification consists of classifying fingerprint images into one out of five categories: whorl (W), right loop (R), left loop (L), arch (A), and tented arch (T). Statistical and structural approaches are followed to this aim.

In particular, the structural approach has received less attention until now. However, a simple visual analysis of the ``structure'' of fingerprint images allows one to see that structural information can be very useful for distinguishing some fingerprint classes (e.g., for distinguishing fingerprints belonging to class A from the ones of class W).

Accordingly, we attempted to ``integrate" flat and structured representations and evaluate the practical benefits of their combination. In particular, we proposed new fingerprint classificationalgorithms based on two machine learning approaches: support vector machines (SVM), and recursive neural networks (RNN).

SVM is a relatively new technique for pattern classification and regression that is well-founded instatistical learning theory. One of the main attractions of using SVM is that they are capable of learning in sparse, high-dimensional spaces with very few training examples. They have been successfully applied to various classification problems and references therein (www.clopinet.com/isabelle/Projects/SVM/applist.html).

A RNN is a connectionist architecture designed for solving the supervised learning problem when the instance space is comprised of labelled graphs. This architecture can exploit structural information inthe data, which, as explained above, may help discriminating between certain classes.

In our work, RNN have been also used to extract a distributed vectorial representation of the relational graph associated with a fingerprint. This vector has been regarded as an additional set of features subsequently used as inputs for the SVM classifier. The system is validate on the NIST database 4. Obtained results have shown that fusion of structural and statistical fingerprint classifiers can reach 90% at 1.8% rejection rate. In particular, the use of RNNs can strongly contribute to better classify fingerprint classes which exhibit a structure more accentuated than other ones (as A and W classes).

People working on this topic:

• Gian Luca Marcialis
• Fabio Roli
• Alessandra Serrau