Updating Labeling of Photographic Segmented Regions using Knowledge Representation Formalism and Dempster-Shafer Theory

This paper discusses the inclusion of explicit symbolic knowledge during the labeling process of segmented images. We intend to update the initial labeling of photographs segmented regions obtained by an existing labeling process. We use the Dempster-Schafer theory of evidence in conjunction with a knowledge representation formalism, namely the Conceptual Graphs. The use of conceptual graphs allows a strong control of the features used to define relevant frames of discernment. The update process uses a set spatial relationships that characterize the image segmented regions, and takes into account learned samples that are not based on segmented regions but on user manual annotations. On a set of more than 350 objects recognized, a significant increase of well-labeled segmented regions is achieved.