A Colour Space Selection Scheme dedicated to Information Retrieval Tasks
Résumé
The choice of a relevant colour space is a crucial step when dealing with image processing tasks (segmentation, graphic recognition...). From this fact, we address in a generic way the following question: What is the best representation space for a computational task on a given image? In this article, a colour space selection system is proposed. From a RGB image, each pixel is projected into a vector composed of 25 colour primaries. This vector is then reduced to a Hybrid Colour Space made up of the three most significant colour primaries. Hence, the paradigm is based on two principles, feature selection methods and the assessment of a representation model. The quality of a colour space is evaluated according to its capability to make colour homogenous and consequently to increase the data separability. Our framework brings an answer about the choice of a meaningful representation space dedicated to image processing applications which rely on colour information. Standard colour spaces are not well designed to process specific images (ie. Medical images, image of documents) so a real need has come up for a dedicated colour model