Segmentation d’images couleur par classification pixellaire et hiérarchie de partitions
Cyril Meurie · HAL (Le Centre pour la Communication Scientifique Directe) · 2005
The first part of this work is devoted to the development of a segmentation strategy of microscopic color images which can contribute to the enrichment of the segmentation stage of an automated microscopy analysis station. We first of all expose several approaches of unsupervised and supervised pixel classifications and show the importance of the color space choice. The complementarity which can exist between classifiers is exploited by proposing a method for combining pixel classifications taking account the number of combined classifiers and the neighbourhood information. Finally we integrate these methods in a morphological approach relying on a color watershed and evaluate the results by an evaluation method adapted to cytology.In a more general context, the second part of this work deals with the segmentation of color images by hierarchy of partitions. We first of all present a new connective criterion and an approach for creating hierarchy of partitions allowing to quickly simplify or segment an image. We then focus on color mathematical morphology and propose a graph approach to determine the infimum and the supremum of color vectors set , this enables a new formulation of the waterfall algorithm. Finally, we define an energy function being which can be used either to automatically determine the best level of a hierarchy of partitions or as a termination criterion for region merging.