A cost-effective fingerprint recognition system for use with low-quality prints and damaged fingertips

Abstract

The development of a robust algorithm allowing good recognition of low-quality fingerprints with inexpensive hardware is investigated. A threshold FFT approach is developed to simultaneously smooth and enhance poor quality images derived from a database of imperfect prints. Features are extracted from the enhanced images using a number of approaches including a novel wedge ring overlay minutia detector that is particularly robust to imperfections. Finally, a number of neural net and statistically based classifiers are evaluated for the recognition task. Results for various combinations of the process are presented and discussed with regard to their utility in such a system.

Introduction

Fingerprints are imprints or impressions of patterns formed by friction ridges of the skin in the fingers and thumbs. The uniqueness of such patterns was recognised by past cultures and their use as a means of personal identification has been in existence as early as the third century B.C. The ability to confirm the identity of crime suspects and to learn the identity of unknown felons from prints left at the scene of a crime is of obvious benefit to society. Thus, it is not surprising that law enforcement agencies are the ones to have made the greatest advances in techniques and equipment in use of fingerprints for positive identification. Closely allied with the art of criminal apprehension, human fingerprints can provide an unequalled method of personal identification for various other applications. These include security or access control systems, banking and credit systems and forensic or surveillance systems.

A shortcoming with many of the existing recognition techniques is that they all require a reasonable standard of input image. Many sophisticated fingerprint enhancement algorithms do exist [1], [2], [3]. However, these techniques are suited for police work, where processing time does not have to be within 1 or 2 s. Thus for a low quality, cheap, fast commercial system there is abundant room for improvement. Coetzee and Botha [4] have investigated this problem with reasonable success. They proposed a recursive line following binarisation and smoothing pre-processing, followed by feature extraction using a Fourier wedge-ring detector which detects minutia which manifest themselves as small deviations from the dominant spatial frequency of the ridges.

A further complication which has not yet been addressed is that of “bad” prints. This refers to fingerprints where the distinguishing patterns are unclear. This may be the result of one of several situations. As one gets older, one's fingerprints tend to wear very quickly and the definition between valleys and ridges significantly deteriorates. Elderly people's ability to replace the cells that build up this definition diminishes, with the result that their fingerprint patterns are difficult to obtain. Manual workers, particularly in rural and mining areas, subject their fingertips to severe punishment, resulting in injuries, scars, calluses and false grooves. All of these contribute to false minutia. Additionally, any system must be robust to people who obtain temporary cuts. Finally, some recognition algorithms are hindered by fingers with significantly different ridge thicknesses both within and between different fingerprints. Combining both low-quality images with low-quality prints, to facilitate a cheap and efficient commercial recognition system, presents a serious challenge. This is the problem that is addressed in this paper.

The noise in fingerprint images is highly correlated and often the statistics of such noises are unknown. Although noise covariance may be estimated or assumed to be Markov spatially, it is not unreasonable to expect that probabilistic noise reduction approaches will not work very well. A better alternative to the statistical approach would be a problem specific technique and thus an efficient fingerprint specific enhancement procedure has been developed. Both binarisation and line thinning are considered and efficiently dealt with. The large amount of false minutia generated may be dealt with in two ways. Xiao and Raafat [5] proposed a post-processing, minutia purification stage to eliminate spurious minutia. This is a costly procedure time-wise and therefore impractical for the problem.

A far simpler, more efficient post-processing procedure is suggested and tested. The other approach, which is the main focus of this paper, is to use the total image (or a robust representation thereof ) to find a match. Thus, a rotationally and translationally invariant description of the entire fingerprint (perhaps with emphasis on minutia) is required for recognition purposes. Standard probabilistic classifiers, neural network classifiers and the newer probabilistic neural network are compared and their performance at the task at hand is evaluated.