Automatic recognition of a person by the use of distinctive physical and behavioral characteristics is called biometrics. Two of the important challenges in biometrics are recognition of identical twins and matching of latent fingerprints to their exemplar prints (rolled or slap prints). The contributions of this dissertation are focused on these two topics.Identical (monozygotic) twins are a result of a single fertilized egg that splits into two cells, each one giving birth to one individual. Identical twins have the same deoxyribonucleic acid (DNA), thus their genotypic features (features influenced by the genetic material) are the same. However, some of their phenotypic features (features influenced by the fetal environment) may be different. Thus, it is essential to determine which biometric traits (either by themselves or in combination with other traits) have the ability to distinguish identical twins and the extent of their ability for this discrimination.The first contribution of this dissertation is an evaluation of the performance of biometric systems in the presence of identical twins for the three most commonly used biometric modalities, namely fingerprint, face and iris. Identical twins are shown to be a challenge to current face recognition systems. On the other hand, fingerprint and iris matching of identical twins show performance comparable to those with unrelated persons. The fusion of different samples from the same modality of a subject (e.g., left and right iris, fingerprints of multiple fingers) yields the best matching performance for identical twins, similar to what has been shown for unrelated persons. Biometric traits can also be used to determine whether two subjects enrolled in a biometric database are identical twins. By using face and iris modalities together, for example, we can correctly identify 80\% of such identical twin pairs, while only 2\% of subject pairs who are not identical twins will be incorrectly considered identical twins.The second contribution of this work is focused on improving latent fingerprint matching performance. Latent fingerprints are partial fingerprint images that typically contain only a small area of friction ridge pattern and large non-linear distortion, are blurred or smudged, and contain complex background noise. Due to these characteristics, latents are a particularly challenging for matching to their mates (reference prints) in a database. Given a latent print in which minutiae have been marked by a human expert (as is the current practice in forensics), we have proposed two approaches to improve the latent matching performance. The first approach consists of enhancing the latent image and fusing the matching score obtained from the enhanced latents with the score based on manually marked minutiae. This approach outperforms a commercial fingerprint matcher on the public latent database NIST SD 27. The second approach consists of developing a latent fingerprint matcher that utilizes minutiae as well as the orientation field information. The proposed matching algorithm outperforms three fingerprint matching algorithms on two different latent fingerprint databases (NIST SD 27 and WVU latent databases).The latent fingerprint identification accuracy generally deteriorates as the size of the fingerprint database grows. To alleviate this problem and to reduce the overall search time of latent matching, we propose to combine various level 1 and level 2 features, including minutia cylinder code, minutia triplets, singular points, orientation field and ridge period, to efficiently filter out a large portion of the reference fingerprint database. The proposed approach outperforms state-of-the-art indexing techniques on the public domain latent database NIST SD27 against a large background database of 267K rolled prints. The experimental results also show that the proposed filtering scheme has the desirable property of attaining improved computational efficiency of latent search (20\% penetration rate) while maintaining the latent matching accuracy.
Read
