Network Representation of 2-D and 3-D Images

Much of the data and knowledge representation in image interpretation needs to be in the form of networks. We propose that a network representation of the grey level changes in an image should be constructed as early as possible during bottom-up processing. We refer to this low level description as the image structure representation. The image is treated as a continuous surface made up of triangular facets (in 2-D) or as a continuous volume of tetrahedrons (in 3-D). The image is completely segmented into nonoverlapping slope districts. Each slope district is anchored between one peak and one pit and (usually) two or more saddle points. Slope districts can easily be grouped in several ways to form more meaningful entities such as edge support regions, ridges, convex corners etc.