Reliable fingerprint recognition

Fingerprints have been used for personal identification for centuries because of their uniqueness and consistency over time. Fingerprint recognition is one of the most popular methods for personal identification due to its high accuracy, cost efficiency and ease of acquisition. Automated fingerprint recognition has the advantages of fast processing and high accuracy, but its performance deeply depends on the quality of the collected fingerprint images. The matching accuracy of current automatic fingerprint recognition systems decreases dramatically when the quality of fingerprint images is poor. For example, a fingerprint image may contain massive noise, cleaves or inks. In these cases, manual fingerprint recognition achieves better matching results than automatic systems. One of the major challenges in fingerprint recognition is how to improve the performance of an automatic fingerprint recognition system in terms of reliability and accuracy, especially for low quality images. The motivation of this research is derived from the raised need for fingerprint recognition techniques with better matching accuracy and reliability. How to improve the accuracy and reliability of an automatic fingerprint recognition system when processing low quality fingerprint images is the major objective of this research work. Because feature extraction and feature matching are two main components in a fingerprint recognition system, the above objective could be restated as: (i) to design reliable and accurate feature extraction techniques suitable for low quality images and (ii) appropriate matching methods or matching metric with high tolerance for image noise and feature extraction errors. In order to achieve the above objectives, effort has been made to improve the matching accuracy of an automatic fingerprint recognition system by introducing the following methods: (i) a fingerprint image pre-processing method in the spatial domain, (ii) two different singular point detection approaches, and (iii) a new matching metric named binarized minutiae block for fingerprint matching. Firstly, we have investigated current fingerprint enhancement techniques. A typical fingerprint enhancement module is composed of an image pre-processing stage and a contextual filtering stage. Traditionally, image pre-processing (or called pixel-wise enhancement) techniques are used to improve the contrast of an image rather than removing noise. In this study, we found that removing noise and improving the image quality in this stage enables the subsequent contextual filtering stage to obtain a better clarity of ridge and valley structure especially for poor quality fingerprint images, particularly suitable for wet and smudged fingerprint images, based on experimental observation. Therefore, we proposed an image pre-processing approach using contrast stretching and power-law transformation techniques to improve the quality of fingerprint images. The metric goodness index (which is used to evaluate the image quality) is used to evaluate this method. The experimental results show that this approach is able to improve the clarity of ridge and valley structures especially for wet and smudged fingerprints. The average goodness index value obtained from the experiment is improved by 14% compared to other reported methods. In addition, it enables the subsequent contextual filtering (e.g. Gabor filtering) stage for better image enhancement results, and ultimately improve the reliability of feature extraction (e.g. minutiae extraction). Secondly, we have investigated feature extraction techniques, especially singular point detection which is a global feature in a fingerprint. The performance of current singular point detection techniques is relatively low for poor quality images (mostly around 90% of correct detection rate, and much lower for Poincaré Index based approaches). As a consequence, it becomes the major bottle neck for fingerprint recognition techniques which rely on singular points, such as reference point based fingerprint global pre-alignment and fingerprint classification. In order to address this issue, we first investigated the popular Poincaré Index based approaches. The Poincaré Index technique highly depends on image quality, and suffers from the problem of a large number of spurious singular points, especially for low quality images. As a consequence, we designed a rule-based post-proccessing technique to validate and remove spurious singular points. The experimental results show that the correct detection rate on average is 89.48% on DB1a and DB2a of Fingerprint Verification Competition (FVC) 2002 datasets. These datasets contain fingerprint images with various quality levels, and are especially suitable for evaluation of fingerprint recognition algorithms. It is around 3% improvement over other reported Poincaré Index based approaches in terms of overall correct detection rate. However, one limitation of the Poincaré Index technique is that it processes data locally while singular points are global features, which are easily influenced by local noise and may cause a number of spurious singular points, especially for low quality images. Therefore, we have proposed a new singular point detection method globally over the whole image, based on the analysis of local ridge orientation maps. In addition, this method is also able to locate a reference point for arch type fingerprints which is useful for fingerprint pre-alignment as a reference point as well as for fingerprint classification. The experimental results show that the correct detection rate on average is 94.05% on the datasets of FVC 2002 DB1a and DB2a. This experimental result is superior to any other reported methods in terms of correct detection rate of singular points. Finally, we have investigated the current fingerprint matching methods, and proposed a new matching metric named binarized minutiae block for fingerprint matching. Current matching methods could be classified as: minutiae based, correlation based, and other non-minutiae based methods. Among these methods, correlation and other non-minutiae based methods have better tolerance to image noise and feature extraction errors than minutiae based methods. However, minutiae based methods have better tolerance to non-linear distortion and obtain better matching results on medium or high quality images. This new metric utilizes the minutiae and its surrounding texture information. Thus, it has high tolerance to image noise and feature extraction errors as well as non-linear distortion. These binarized minutiae blocks are normalized to the same minutiae direction for easy comparison. Then, the local similarities are calculated by the dissimilarities between each pair of binarized minutiae blocks. In addition, four global similarity calculation methods are designed and implemented using this matching metric. The experimental results show that this method achieves overall matching accuracy of 98.24%, 97.87% and 98.19% on the datasets FVC2002 DB1a, DB2a and FVC2006 DB2a. As a consequence, the results suggest that using binarized minutiae blocks is an alternative way to obtain accurate and reliable matching results other than correlation based (grey scale texture information), minutiae based and other non-minutiae based methods. Compared to other state-of-the-art matching methods, this metric achieves better experimental results in terms of matching accuracy than most reported matching methods on the same testing databases. In conclusion, this thesis focuses on the research of how to improve the overall matching accuracy of a fingerprint recognition system even for low quality images. Several methods have been developed to achieve this research objective. The experimental results show that these proposed fingerprint recognition techniques are able to improve the recognition accuracy significantly.