Segmentation of Natural Images Using Scale-Space Representations: A Linear and a Non-Linear Approach

In general purpose computer vision systems, unsupervised image analysis is mandatory in order to achieve an automatic operation. In this paper a different approach to image segmentation for natural scenes is presented. Scale-Space representation is used to extract the structure from meaningful objects in the image. Two different scale-spaces are analysed in the paper. On one hand Isotropic Diffusion (linear scale-space) is presented as the basis for an uncommitted front end, not relying on any special feature of the image. On the other hand the Total Variation Diffusion (non-linear scale-space) which makes a special emphasis on edges is also analysed. A hierarchical decomposition of the image is performed on the basis of the special characteristics of each scale-space. Iso-intensity paths will be tracked in the case of linear scale-space, whereas in the case of non-linear scale-space the evolution of level sets through scale will be tracked. In the framework of linear scale-space, the use of additional information to improve the robustness in the structure extraction is introduced. Appart from the set of several diffused versions of the image, a representation of edges through scale is included to supervise the generation of the hierarchical tree that represents the image.