Applications of linear weight neural networks to fingerprint recognition

Fingerprints form an important aspect of evidence in criminal investigations in modern police work. However, the task of searching for a match from a scene-of-crime image (a mark or latent) to the files of prints taken from previous convicts can be labour-intensive. The new approach described in this paper uses a localised ridge direction determination which is generated by applying anisotropic filters; consequently this ridge flow estimate is largely immune to ridge degradations. The surface produced by this processing is then input to a Volterra classifier. The same classifier has had weights trained to find the fingerprint core delta and pattern classification, by using human-marked images. The choice of the correct neural net architecture was vital, and several approaches were tested. An approach which was linear in the weights (Volterra radial basis functions and so forth) was chosen. The approach has been in operational use for large-database criminal fingerprint systems with the UK Police for two years and has recently been adapted for biometric fingerprint verification. This application makes use of the robustness of the algorithm to produce a low-cost system using low-quality imaging optics and camera. The neural approach shows particular robustness and allows further automation of the fingerprint encoding process. This in turn allows more information to be extracted for each point of interest and consequently higher matching performance.