Observatorio de I+D+i UPM

Memorias de investigación
Nuevas contribuciones en las representaciones sobrecompletas de imágenes inspiradas en la arquitectura funcional de la corteza visual primaria
Áreas de investigación
  • Industria electrónica
The present thesis aims at investigating parallelisms between the functional architecture of primary visual areas and image processing methods. A first objective is to refine existing models of biological vision on the base of information theory statements and a second is to develop original solutions for image processing inspired from natural vision. The available data on visual systems contains physiological and psychophysical studies, Gestalt psychology and statistics on natural images. The thesis is mostly centered in overcomplete representations (i.e. representations increasing the dimensionality of the data) for multiple reasons. First because they allow to overcome existing drawbacks of critically sampled transforms, second because biological vision models appear overcomplete and third because building eficient overcomplete representations raises challenging and actual mathematical problems, in particular the problem of sparse approximation. The thesis proposes first a self-invertible log-Gabor wavelet transformation inspired from the receptive field and multiresolution arrangement of the simple cells in the primary visual cortex (V1). This transform shows promising abilities for noise elimination. Second, interactions observed between V1 cells consisting in lateral inhibition and in facilitation between aligned cells are shown efficient for extracting edges of natural images. As a third point, the redundancy introduced by the overcompleteness is reduced by a dedicated sparse approximation algorithm which builds a sparse representation of the images based on their edge content. For an additional decorrelation of the image information and for improving the image compression performances, edges arranged along continuous contours are coded in a predictive manner through chains of coefficients. This offers then an efficient representation of contours. Fourth, a study on contour completion using the tensor voting framework based on Gestalt psychology is presented. There, the use of iterations and of the curvature information allow to improve the robustness and the perceptual quality of the existing method.
Tipo de Tesis
Sobresaliente cum laude
Esta actividad pertenece a memorias de investigación
  • Autor: Sylvain Fischer
  • Director: Andres de Santos Lleo (UPM)
  • Director: Gabriel Cristóbal (Instituto de Optica, CSIC)
Grupos de investigación, Departamentos, Centros e Institutos de I+D+i relacionados
  • Creador: Grupo de Investigación: Tecnología de imágenes biomédicas
  • Departamento: Ingeniería Electrónica
S2i 2022 Observatorio de investigación @ UPM con la colaboración del Consejo Social UPM
Cofinanciación del MINECO en el marco del Programa INNCIDE 2011 (OTR-2011-0236)
Cofinanciación del MINECO en el marco del Programa INNPACTO (IPT-020000-2010-22)