Palmprint recognition for forensic applications=Reconocimiento de huella palmar para aplicaciones forenses

• Autores: Ruifang Wang
• Directores de la Tesis: Daniel Ramos Castro (dir. tes.), Julián Fiérrez Aguilar (dir. tes.)
• Lectura: En la Universidad Autónoma de Madrid ( España ) en 2013
• Idioma: inglés
• Tribunal Calificador de la Tesis: Javier Ortega García (presid.), Luis Salgado Álvarez de Sotomayor (secret.), Christiane Batista de Paulo Lobato (voc.), Miguel Ángel Ferrer Ballester (voc.), Raymond Veldhuis (voc.)
• Materias:
  • Física
    • Electrónica
• Enlaces
  • Tesis en acceso abierto en: Biblos-e Archivo
• Resumen

  • Robust and efficient automatic forensic palmprint recognition remains a big challenge, mainly due to the large number of creases and large non-linear distortion in high-resolution palmprints, and the large storage and computation capabilities required for processing and matching palmprints. Compared to forensic fingerprint recognition, the techniques of forensic palmprint recognition have recently received attention while palmprint evidence is increasingly important according to surveys by law enforcement agencies.

    This Ph.D. Thesis is focused on building robust and efficient palmprint recognition systems for forensic applications. Due to the fact that few systems aimed at forensic palmprint recognition have been developed, deeper research is needed in this field in order to improve robustness, accuracy and computational efficiency of automated forensic palmprint recognition technologies, aimed at realistic forensic scenarios. Moreover, in order to foster the use of palmprint recognition systems in forensic applications, advanced algorithms with multi-system and multi-region fusion, and a proper evidence evaluation framework, will constitute a significant advance.

    In this context, the present Ph.D. Thesis pretends to contribute with novel methods to face the challenges of forensic palmprint recognition. In particular, the Thesis takes advantage of a better understanding of large non-linear distortion problem in forensic palmprint recognition, proposes novel algorithms dealing with the rich types of features in a high-resolution palmprint image and fusion of palm regions, and performs evidence evaluation using the similarity scores, i..e, the output of forensic palmprint recognition systems.

    This Ph.D. Thesis devotes to the so-called field of forensic biometrics, a new area arising from the fields of biometrics and forensic science. Therefore, it requires background of both fields, as well as their interaction. We start with this point and give an introduction of basic concepts involved in forensic palmprint recognition and evaluation. To have a general view of the tasks, we introduce the focuses of forensic science, palmprint recognition, fusion in biometrics and evidence evaluation.

    To better understand the main problems of forensic palmprint recognition, we review related works in this field. Some state-of-the-art algorithms in high-resolution palmprint recognition can be applied to forensic applications, which are mainly based on minutiae features in high-resolution palmprint images. As most forensic fingerprint recognition systems also follow minutiae-based matching, novel techniques of minutiae feature representation for fingerprints can also be applied to forensic palmprint recognition. According to the practices in the field of forensic biometrics, approaches of likelihood ratio based evidence evaluation using biometric systems, such as forensic speaker and fingerprint recognition systems, can also be used for palmprint evidence evaluation.

    A reliable experimental framework is essential for the development of the methods proposed in this Thesis. Following the principle practices in the field of forensic biometrics, we include protocols and measures for both performance evaluation of biometric systems and likelihood ratio based evidence evaluation methods, in order to comparatively assess the experimental parts in this Ph.D. Thesis. The experimental framework also includes a description of the databases of palmprints and fingerprints we use in this Thesis. For both palmprints and fingerprints, public databases and unpublished forensic databases are used. In the case of public databases, THUPALMLAB for palmprints, NIST SD27 and FVC 2002 for fingerprints are used. As unpublished forensic databases from criminal investigations, a forensic fingerprint database collected by Netherlands Forensic Institute and forensic palmprint databases collected by Beijing Institute of Criminal Technology in China are used.

    The first forensic palmprint recognition system we propose in this Thesis is based on radial triangulation representation for minutiae in palmprints. The proposed complete forensic palmprint recognition system based on radial triangulation includes both novel forensic palmprint feature extraction and comparison, namely, the MinutiaLine extractor and RT comparator. Moreover, a combined global feature comparison component is implemented, in which weights of centroids of radial triangulation structures and principal line energy are both considered for finer global comparison. The MinutiaLine extractor outperforms a commercial extractor MegaMatcher 4.0 with much less spurious minutiae extracted and performs much faster than the MinutiaCode extractor. The composed MinutiaLine+RT system performs better than the MinutiaCode-based system regarding both accuracy and computational efficiency. Moreover, the proposed method of combined global comparison outperforms centroid-based and principal line-based global comparison methods.

    We then present the second forensic palmprint recognition system we propose in this Thesis which is based on weighted complex spectral minutiae representation (Weighted-SMC). We first conduct a meaningful study of distortion assessment at feature level for fingerprints and palmprints in forensic scenarios, which can guide the design of forensic fingerprint/palmprint recognition systems. Inspired by the study, a novel Weighted-SMC comparator is proposed and implemented as the key component of the recognition system. With applications first to forensic fingerprint comparison and then to forensic palmprint comparison, the Weighted-SMC comparator works in a comparable way for fingerprints compared to the results of the public evaluation, i.e., NIST ELFT-EFS Evaluation, and outperforms RT-based comparator for palmpritns.

    Aimed to improve the performance of forensic palmprint recognition systems by information fusion, fusion scheme for forensic palmprint recognition is studied. Using the two forensic palmprint recognition systems we propose in this Thesis, the evaluation of multi-algorithm fusion at score level is conducted as one part of our fusion scheme for forensic palmprint recognition. Another part of the fusion scheme is implemented by anthropologically-inspired regional fusion for high-resolution palmprint recognition in full-to-full matching mode. Existing regional fusion proposals in the literature are mainly based on simply-shaped or equally-divided sectors. To achieve more meaningful regional fusion, our proposal is anthropologically motivated, dividing the palm in three different regions according to anatomical constrains, i.e., utilizing datum points located on principal lines of a palmprint. This is more adapted to the view of a forensic practitioner, and is relevant as a study of the distribution of discriminating information in the different regions of a palm, namely, regional discriminability. To obtain palm regions based on datum points, both manual segmentation and automatic segmentation are implemented. The evaluation of regional fusion at score level is also conducted. A significant improvement by the proposed regional fusion is achieved for high-resolution palmprint recognition in full-to-full matching mode which is also important in forensic applications such as preventing identity spoofing.

    The research work described in this Dissertation has led to novel contributions which include two new methods of forensic palmprint comparison and new methods on automatic region segmentation for high-resolution palmprints and distortion assessment in forensic scenarios, namely: i) local comparison based on radial triangulation according to point pattern comparison by relaxation; ii) forensic fingerprint/palmprint comparison based on weighted complex spectral minutiae representation (Weighted-SMC); iii) automatic region segmentation based on convex hull comparison for high-resolution palmprint recognition; iv) distortion assessment using minutiae window. Moreover, a framework for high-resolution/forensic palmprint recognition system based on anthropologically-inspired regional fusion has been proposed first time. Besides, some literature reviews has been derived from this Dissertation.