Fingerprint recognition has been successfully used in law enforcement and forensics to identify suspects and victims for over a century. Recent advances in automated fingerprint identification systems (AFIS), coupled with the growing need for reliable person recognition, have resulted in an increased deployment of AFIS in broad applications such as border control, employment background checks, secure facility access, and user authentication in laptops and mobile devices. Despite the success of fingerprint recognition technique in many large-scale and diverse person identification applications, several challenging issues in fingerprint recognition still need to be addressed. First, the persistence and uniqueness of fingerprints---the fundamental premises for fingerprint recognition---remain as presumptions rather than facts with solid scientific underpinnings. Although some studies have addressed the uniqueness of fingerprints, there has been no systematic study reported on the persistence of fingerprints. Given a large longitudinal fingerprint database, we have analyzed it with a multilevel statistical model to assess the impact of time interval between a genuine fingerprint pair on the corresponding match score and identification decision. Second, an appropriate mathematical model for fingerprint orientation field is necessary in addressing a number of challenging problems, including feature extraction (e.g., orientation field and minutiae) from noisy or partial fingerprints, detecting abnormality in fingerprint patterns, and representing fingerprints in a compact form. To this end, we have developed a global orientation field model in the form of ordinary differential equations which is constrained by the number of singular points in a fingerprint. The model represents the global orientation field with a small number of polynomial terms and outputs fingerprint-like orientation fields with the specific number of singular points after fitting an input pattern. We use this model to check the fingerprint-ness of the input image and ensure the integrity of exemplar fingerprint databases. Third, given that automatic latent fingerprint matching is difficult due to poor quality of latent fingerprints found at crime scenes, we have developed a semi-automatic latent enhancement method. The proposed algorithm is based on robust orientation field estimation that is able to (i) reduce human intervention in feature markup and (ii) improve automatic feature extraction and matching accuracy of latents. Fourth, fingerprint obfuscation or alteration is a type of attack on AFIS that is of concern to law enforcement and border crossing agencies. We show that the fingerprint matching accuracy can greatly deteriorate when the altered fingerprints are presented to AFIS. To address this deficiency of AFIS, we have (i) analyzed the types of fingerprint obfuscation, (ii) developed a detection algorithm for altered fingerprints by measuring the abnormality in orientation field and minutiae distribution, and (iii) proposed a restoration and matching algorithm to possibly link an altered fingerprint to its pre-altered mate. By addressing these contemporary problems, this dissertation has advanced our understanding of fingerprint recognition and enabled development of robust and accurate fingerprint recognition algorithms.
Read
