LABEL MAPS FUSION FOR THE MARGINAL SEGMENTATION OF MULTI-COMPONENT IMAGES

Abstract

In this paper, we propose a new technique for merging the label maps obtained by the marginal segmentation of a multi-component image. In the marginal segmentation, each component of the multi component image is independently segmented by labeling the pixels of the same class with the same label. Therefore the number of label maps corresponds to the number of components in the image. It is then necessary to merge them in order to have a single label map, i.e. a single segmented image. In the most merging techniques, the compatibility links between these maps are performed at prior by making the correspondences between their labels. However the various components are segmented and labeled independently, label maps are considered as independent sources. It is then difficult to establish the relationship compatibilities between labels. The method we propose does not assume any compatibility link at prior. The label maps are combined by superposition. Unfortunately, an over-segmentation is produced. To cope with this problem, the insignificant regions and classes are eliminated. Finally, classes are grouped by using hierarchical agglomerative clustering algorithm. Tests performed on color and satellite images show the effectiveness of this method and its superiority compared to the vector segmentation. Self Organizing Maps algorithm is used during the segmentation process in the both marginal and vector segmentations.