Fingerprint recognition for forensic applications
- Autores: Ram Prassad Krishnamoorthy
- Directores de la Tesis: Daniel Ramos Castro (dir. tes.), Julián Fiérrez Aguilar (dir. tes.)
- Lectura: En la Universidad Autónoma de Madrid ( España ) en 2015
- Idioma: español
- Tribunal Calificador de la Tesis: Javier Ortega García (presid.), Roberto Paredes Palacios (secret.), Josef Bigun (voc.), Juan Carlos San Miguel Avedillo (voc.), Raymond Nicolaas Johan Veldhuis (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: Biblos-e Archivo
- Resumen
Forensic latent fingerprint recognition has taken a significant transition in recent decades from a fully manual identification procedure to the incorporation of Automated Fingerprint Identification Systems (AFIS) by law enforcement agencies to identify suspects. Latent fingerprints are those fingerprints that are revealed using chemical or optical processing and collected from a crime scene by specialists trained in forensic sciences. These latent fingerprints are then compared to reference fingerprints stored in the AFIS database called tenprints or exemplars. However, the latent fingerprints thus obtained from crime scenes are generally of very poor quality. They can be highly partial in nature, non-linearly deformed, smudgy, and with overlapped fingerprints. These characteristics of latent fingerprints introduce many challenges in employing a fully automatic latent fingerprint matching.
This PhD Thesis is focused on solving some of the major challenges in automated latent fingerprint identification. In particular, the Thesis focuses on improving the performance accuracies of minutiae-based fingerprint matchers for latent fingerprints by exploring the problem of partial fingerprint to full fingerprint matching. Further, we also develop an evidence evaluation model for the matched fingerprints based on likelihood ratio from the similarity scores generated by the minutiae-based matchers. More specifically, the main contributions of the Thesis can be divided in three blocks.
First, most fingerprint matching algorithms in general assume approximately the same size of the minutiae set, the set of key features of a fingerprint based on which comparisons can be made, between the query and the reference minutiae for good identification accuracy. In practice, however, it is frequent that the size of latent minutiae set is very small compared to the size of tenprint minutiae sets, as a result of which matching performance is hampered. To make the best use of minutiae-based matchers, it is advantageous to know the location of partial fingerprint minutiae pattern in the full fingerprint minutiae pattern, thereby reducing the minutiae search space to improve the matching performance. However, existing minutiae based alignment techniques are not well adapted to use in partial fingerprint alignment. An image-based registration is also not feasible due to the poor quality of latent fingerprints. In the first part of this Thesis, we focus on the problem of aligning a partial fingerprint against a full fingerprint, especially of poor quality latents. Instead of minutiae, we used orientation fields (OF) to perform the alignment. We reduce fingerprint images to orientation images, and we look at the alignment problem as registering the partial fingerprint orientation image into the full fingerprint orientation image. To achieve this task, we develop a new correlation-based hierarchical registration method for orientation images to register a partial fingerprint in a full fingerprint. The OF representing the flow of ridges is a relatively stable global feature of fingerprint images. We experimentally demonstrate a significant improvement in the rank identification accuracies for minutiae-based matchers by incorporating our registration algorithm to reduce the search space of minutiae in full fingerprints. We also demonstrate the usefulness of our proposed method as a fully automatic tool.
Second, AFIS use only a limited subset of the types of discriminatory features which can be automatically extracted from the fingerprint images using a feature extraction algorithm. On the other hand, forensic examiners use a richer set of features during manual comparison as compared to AFIS comparisons. This could be a possible reason why manual comparisons outperform AFIS comparisons. The features not currently used by commercial AFIS are generally termed as Extended Feature Sets (EFS). Many commercial minutiae-based matchers do not use EFS. They mostly use only two prominent ridge characteristics namely ridge-endings and bifurcations. To use EFS in automated systems, reliable feature extraction algorithms are mandatory. In the second part of Dissertation, we focus on the problem of using EFS in a typical minutiae-based matcher. A realistic database from forensic fingerprint casework consisting of rare minutiae features was obtained from the Spanish law enforcement agency, Guardia Civil. We propose a method to improve the identification accuracy of minutiae-based matchers for partial latent fingerprints by incorporating reliably extracted rare minutiae features. Our proposed algorithm modifies the similarity scores of minutiae-based matchers based on the presence of rare minutia features like fragments, enclosures, dots, interruptions, etc. These rare features are used to automatically estimate an affine function that transforms the latent minutiae set to the tenprint minutiae set, generating a fitting error which is then used to adjust the baseline minutiae-based matching score. We experimentally demonstrate significant improvement in the rank identification accuracies of minutiae-based matchers when their similarity scores are modified in this way.
Third, the uniqueness of a fingerprint is not an established fact but only an empirical observation. There is a widespread concern about the scientific basis underlying the individuality of fingerprints, especially when using them in the court of law. Many individualization models for fingerprints have been proposed in the research literature. However, there is no scientific framework in use at the criminal justice system to characterize the uncertainty involved in the friction ridge analysis methodology, as well as to express the strength of opinion of the forensic examiner quantitatively. Such a requirement has been articulated in several influential reports like the National Research Council 2009 report ``Strengthening Forensic Science in the United States: A Path Forward''. The new paradigm coming forward in this regard avoids hard identification decisions by considering evidence reporting methods that incorporate uncertainty and statistics. Among all the methods for evidence evaluation, the likelihood ratio has shown much promise and is receiving greater attention. Using the technique we developed for improving the rank identification accuracies of minutiae-based matchers by incorporating rare minutiae features outlined above, we build a robust likelihood ratio evidence evaluation model for individualization.
In summary, in this Dissertation we address three key challenges for automated latent fingerprint matching: 1) partial fingerprint registration using Orientation Fields, 2) use of Extended Feature Sets, and 3)development of a robust evidence evaluation tool.
