HDR defense Damien Muselet"Models and data for the extraction of physical, invariant, or discriminative information from color images."

"Models and data for the extraction of physical, invariant, or discriminative information from color images"

abstract

A color image is made of of pixels characterized by three color components providing information on a quantity of red, green and blue. Thus, the spatial distribution of color vectors in the image constitute the only available information allowing to solve classical computer vision tasks such as image classification, object detection or semantic segmentation. Therefore, it is necessary to develop solutions that enable the extraction of relevant local information from the image and, if needed, to merge these characteristics into a global semantic descriptor.

In this context, I have exploited the spatial distribution of color vectors that constitute an image with three main objectives in mind:

• Estimation of physical information related to the different elements involved in image formation (light source, surfaces, sensors...).
• Extraction of local descriptors or color coordinates that provide the best possible balance between invariance and discriminative power.
• Selection and fusion of these local descriptors into a global descriptor relevant to a given task.

Pursuing these objectives required the exploitation and development of physical models of color formation as well as machine learning tools.

committee

• Tougne Laure (Univ. Lyon 2) Rapportrice
• Macaire Ludovic (Univ. Lille 1) Rapporteur
• Burie Jean-Christophe (Univ. La Rochelle) Rapporteur
• Gouiffes Michèle (Univ. Paris-Saclay) Examinatrice
• Trémeau Alain (Univ. Saint-Etienne) Tuteur