Automatic fingerprint classification scheme using template matching with new set of singular point-based features

Abstract

Fingerprint classification is a technique used to assign fingerprints into five established classes namely Whorl, Left loop, Right loop, Arch and Tented Arch based on their ridge structures and singular points’ trait. Although some progresses have been made thus far to improve accuracy rates, problem arises from ambiguous fingerprints is far from over, especially in large intra-class and small inter-class variations. Poor quality images including blur, dry, wet, low-contrast, cut, scarred and smudgy, are equally challenging. Thus, this thesis proposes a new classification technique based on template matching using fingerprint salient features as a matching tool. Basically, the methodology covers five main phases: enhancement, segmentation, orientation field estimation, singular point detection and classification. In the first phase, it begins with greyscale normalization, followed by histogram equalization, binarization, skeletonization and ends with image fusion, which eventually produces high quality images with clear ridge flows. Then, at the beginning of the second phase, the image is partitioned into 16x16 pixels blocks - for each block, local threshold is calculated using its mean, variance and coherence. This threshold is then used to extract a foreground. Later, the foreground is enhanced using a newly developed filling-in-the-gap process. As for the third phase, a new mask called Epicycloid filter is applied on the foreground to create true-angle orientation fields. They are then grouped together to form four distinct homogenous regions using a region growing technique. In the fourth phase, the homogenous areas are first converted into character-based regions. Next, a set of rules is applied on them to extract singular points. Lastly, at the classification phase, basing on singular points’ occurrence and location along to a symmetric axis, a new set of fingerprint features is created. Subsequently, a set of five templates in which each one of them represents a specific true class is generated. Finally, classification is performed by calculating a similarity between the query fingerprint image and the template images using x2 distance measure. The performance of the current method is evaluated in terms of accuracy using all 27,000 fingerprint images acquired from The National Institute of Standard and Technology (NIST) Special Database 14, which is de facto dataset for development and testing of fingerprint classification systems. The experimental results are very encouraging with accuracy rate of 93.05% that markedly outpaced the renowned researchers’ latest works.