Mathematical Morpholgy Applied to the Analysis of the lighting Conditions of Color images


Abstract

In this thesis some color image analysis tools are presented aiming at extracting pertinent information on the lighting conditions in which the picture was taken.  Trough these tools, we seek to understand the semantic content of a picture by the way it is lit. This knowledge is useful in several imaging domains such as augmented reality, motion picture post-production, image indexation and pattern recognition.

Light's intrinsic information is not directly available in the digital image which is the result of the integration and the digitalization of the incident spectral flux with the scene's spectral reflectances.  Thus, some semantic objects of interest according to our problem are identified and the tools needed to analyze them are implemented. To do so, physical lighting models are used in order to describe the light reflection behavior and to understand how it modifies the image data.

Firstly, a photometric approach to the analysis of the lighting conditions is presented. It is built around a shadow detection tool. The shadow information content allows the measurement of a luminance contrast over the whole image, giving a ratio between the energy of the direct light source and the ambient light source. In order to increase the accuracy of the analysis, a sky detection tool is developed. It enables the extraction of the meteorological conditions prevailing when the picture was taken. Indeed, the lighting conditions vary greatly from a clear sky to a covered one.

Secondly, we present a method for detecting the illuminant's chrominance. This tool takes advantage of the chromatic convergence which is based on the dichromatic model. The convergence observed on non-homogeneous surfaces is used in order to identify the illuminant's chrominance. The problems inherent to the existing techniques, related to the statistical nature of the processing implemented, are bypassed by using the morphological color segmentation. It allows segmenting the image in regions homogeneous in color and luminance; each region being made of a single spectral reflectance. A filtering step is also implemented in order to filter out the regions which do not verify the assumptions made by the dichromatic model. Finally, the chromatic convergences computed on each region are projected on the chromatic diagram. The most probable intersection of the set of all straits with the locus of the planckian radiators is the point of convergence. This point gives the coordinates of the illuminant's chrominance.

 


 Key words

Lighting conditions, Color image analysis, Dichromatic reflection model, Shadow detection, Sky detection, Dichromatic color constancy, Color morphological segmentation
 



 
 
Valery Risson

Last modified: Tue Jan 29 16:29:15 CET 2002Page d'accueil du CMM