A scene interpretation system that solves multiplicity of scene description

It is desirable to arrive at the scene description at a higher level, by grouping such image features as edges and region boundaries to generalize the scene interpretation or to recognize the object with a structure. Often, however, the feature extraction and the grouping are not unique, conflicting, or contain errors. Here, this problem is called the multiplicity of scene description. The basic framework is presented for constructing a system to solve this problem, which is composed of the following two processes: (1) the image features and the spatial relationships are considered as hypothesis and are managed in ATMS. The conflicting scene descriptions are represented by the multiple context; and (2) the bottom-up processing is realized by the rule base. The rule base is composed of the hypothetic knowledge permitting contained errors and the constraining knowledge declaring the relations among alternative data. Based on this framework, the following system is realized. The feature extraction and structurization are executed for the stereo color gray-level images. A subclass of generalized plane composing an orthogonal hexahedron are derived as the three-dimensional description elements for the scene. Then the scene is interpreted by the matching to the model. In the proposed system, the scene description and interpretation are derived as more than one maximal set composed of the hypothesis together with its derivation, which can account for the largest number of observed edges and region boundaries. Some examples of processing also are shown.