Fingerprint classification : a BI-resolution approach to singular point extraction

Abstract

Fingerprint has been used as a biometric feature for security reasons for a long time. Studies have shown that fingerprint is unique to each and every person, even between two identical twins. Nevertheless, each fingerprint structure has a noticeable pattern, which is easily spotted by the naked eye. This flow-like pattern can be classified into five main categories: Arch, Tented Arch, Left Loop, Right Loop and Whorl. They can be used to classify fingerprints and improve the performance of fingerprint recognition. Various fingerprint classification scheme using singular points have been known in literature. Extracting singular points have been found to be an error prone process, depending on the quality of the fingerprint image. This thesis presents a singular point extraction method using two layers of information, without pre-processing of the image. The directional image is computed with Gaussian smoothed squared gradients pixel wise directly from the greyscale image. Two resolutions of the directional image are taken. The coarser level is used to estimate the position of singular points. The algorithm will only search for singular points in the finer level in a particular region if and only if there is a hit. Since the fingerprint is not segmented apriori, the algorithm makes use of the strength of the directional image as region of interest. The algorithm is tested with the NIST 4 fingerprint database. The result of this algorithm is very promising.