MMA-based region localisation for iconic image representation and transmission

A novel method for identifying regions of interest and producing iconic caricatures is presented. The generation of these caricatures is designed for a context-based form of progressive image transmission. The identification of regions of interest uses shape information provided by multiscale medial axis (MMA). The multiscale medial axis and its associated width information are used to initialise boundary and region localisation. This approach avoids the need for prior knowledge of composition to segment and label regions of an image. Results comparing iconic images generated using a previously published method are presented. These results show improved ability to detect low contrast regions and delineation appropriate to the context to give a good caricature. These results confirm that robust and reliable iconic representation can be generated without prior knowledge of image content. The method described here also has potential for extension to 3D image data sets.