Métodos para extração de atributos em imagens de impressão digital.

Abstract

Among the various characteristics of individuals that can be used in a biometric system, fingerprints have been widely used as they enable high accuracy and require low-cost equipment. However, fingerprint recognition is still a problem with gaps for improvement on the false acceptance and false rejection errors present in state-of-the-art algorithms. The reliability of these algorithms depends on the quality of the fingerprint image and the information extracted to perform recognition. In this context, the present thesis proposes new methods for attribute extraction using innovative segmentation, enhancement, and quality definition techniques. We aim to reduce error rates and achieve competitive state-of-the-art results. This thesis presents a new algorithm for evaluation of the global quality from a fingerprint image, which contributes to the false positive reduction and decreases the error rates. The results showed that 10% of the worst quality images are responsible for more than 60% of the errors. Also, a new algorithm for delimiting the region of interest of the fingerprint was developed, outperforming competing works with an average increase of 5.6% on accuracy. Finally, we propose improvements in the fingerprint enhancement process inspired by the following approaches found in the literature: contrast adjustment, Gabor filters, and frequency domain filtering. The methods were evaluated in respect to the error rates obtained by an algorithm for fingerprint matching using the Fingerprint Verification Competition - (FVC 2000, 2002, 2004 and 2006) and FVC OnGoing databases. The methods presented competitive results in comparison to other approaches on the same databases.