Analysis of image processing, machine learning and information fusion techniques for contact-less hand biometrics

• Autores: Belén Ríos Sánchez
• Directores de la Tesis: Carmen Sánchez Avila (dir. tes.) , Hiroshi Ishikawa (codir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2017
• Idioma: español
• Tribunal Calificador de la Tesis: Luis Alfonso Hernández Gómez (presid.) , Asunción Santamaría Galdón (secret.) , Raúl Sánchez Reillo (voc.) , Richard F. Gunst (voc.) , Enrique Cabello Pardos (voc.)
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • In this thesis different techniques of image processing, machine learning and information fusion have been analysed in relation to their applicability to contact-less hand biometrics. To this end, a modular and configurable system that explodes the multimodal nature of the hand to increase its robustness and accuracy have been designed, implemented and evaluated. Given the fact that different applications have different accuracy and time performance needs, the evaluation is aimed to provide a fair comparative of methods under different environmental conditions that helps to adapt the system to the specific requirements of a concrete final application.

    A correct hand segmentation is necessary to extract reliable and invariant biometric features. For this reason, a comparative of different segmentation methods that include well-known methods such as global thresholding and graph cuts as well as a novelty flooding-based method which combines different image-based segmentation approaches. These methods have been compared using diverse datasets of images which cover a wide spectrum of capturing conditions.

    On the other hand, a comprehensive evaluation of different palmprint feature extraction methods comprising Gabor and Sobel filters, Local Binary Patterns, Local Derivative Patterns and Curvelets has been carried out. Different parameter configurations have also been tested with the aim of finding out which arrangement provides better results for each method. In addition to palmprint, also hand geometry features have been extracted. This evaluation includes also two different feature matching approaches: distance-based and Support Vector Machines.

    In addition, it has also been evaluated the feasibility of combining different feature extraction methods to yield into a more precise and robust multimodal solution. Two different levels for fusing the biometric information have been compared: score-level and feature-level.

    Finally, an evaluation methodology that allows for a fair comparison between different methods has been proposed. In particular, an evaluation protocol is offered with the aim of not only obtaining an extensive evaluation of the complete system under different environmental conditions, and testing multiple combinations of methods for each module, but also providing a basis against which to compare future research.