Fingerprint classification techniques

Today, most biometrics research groups are tackling the challenging problem of an automatic fingerprint identification system (AFIS) using large databases. Since AFIS dedicates most of its processing time to searching for the best-matched fingerprint, searching over the entire fingerprint database is rather inefficient. It is proposed that the database be divided into sub-databases, each containing only fingerprints of the same pattern or class. Fingerprint classification is then an important first step in directing the search only to the appropriate sub-database, thus reducing the extent of searching of the large database. The main objective of this thesis is to propose a classification technique to reliably classify a fingerprint into one of six well-known classes: plain arch, tented arch, right loop, left loop, whorl and twin loop. The fingerprint classification technique proposed in this thesis has achieved good results owing to the improvement in a number of processing steps the author has proposed for the enhancement of fingerprints, the determination of singular points and their associated principal axes, and the rule-based classification algorithm. The directional bandpass Gabor filter-bank approach is one of the most effective and mathematically elegant techniques to date for fingerprint image enhancement. The filter output, however, is very sensitive to the ridge orientation and frequency that the filter is tuned to, and also to the spatial parameters of the Gaussian envelope. Unfortunately, filtering of a fingerprint image with an adaptive two-dimensional Gabor filter bank is computationally expensive because ridge orientation and frequency vary significantly throughout the fingerprint. In this thesis we propose to use an array of 8x4 two-dimensional Gabor filters tuned to eight directions and four ridge frequencies. Filtered fingerprint images at any combination of local ridge orientation and frequency can be calculated using a 2-D interpolation algorithm. The proposed technique produces a better quality of image than current Gabor-based techniques. The results are compared using a goodness index measure of the reliability of the automatic minutiae detection. The accuracy of the location of singular points on a fingerprint is important for minutiae matching alignment and is also essential for the Poincare index to correctly determine the type of singular points. In this thesis, we present a novel yet simple and accurate technique for the automatic determination of singular points. The technique offers a double-resolution estimation of the ridge orientation on a 4x4 pixels quincunx grid and quantises ridge orientations into six codes called ridge flow codes. Singular regions are defined as where all six codes exist. A singular point within a singular region can then be quickly determined by locating the point where the variance of local ridge orientation is at its maximum. The Poincare indices of these singular points are used to determine their type: ordinary, delta, core or double-core. The distribution and type of the singular points, together with their associated principal axes, are then used to classify a fingerprint into one of six well-known classes or patterns. This thesis proposes a rule-based algorithm for classifying fingerprints into one of six well-known classes. The rules are formed using the relative locations and types of singular points and the relative directions of their associated principal axes. The reliable and fast classification algorithm is made possible by a simple but effective combination of ridge flow-code technique and orientation variance calculation in the determination of singular points and principal axes. The Poincare indices of these singular points are used to determine their type: ordinary, delta, core or double-core. For a test sample of some 150 fingerprints, the correct classification rate of the proposed algorithm was found to be better than 90%.