Abstract
Human's image understanding process can be roughly divided into the following two kinds of processes; a perceptual grouping phase and a hypothesis recognition phase. In the former, uncertainties mainly come from the quality of the image itself, while in the latter phase some "dynamic" uncertainties are brought about during mutual interactions between the pre-existing knowledge and the arriving data. The image dealt with in this paper is the one called as a "lineament map", that represents a distribution of linear features on the surface of the earth that are observed in the remotely sensed satellite imagery, and the final goal is to identify one of the prototypical geological hypotheses for the input image. For this purpose, we at first propose a way of applying fuzzy clustering methods in an incremental fashion for extracting fuzzy features out of the input image. Then, to model a mutually interacting process between the data and the knowledge, we propose structured neural networks that consist of three different Hopfield-type networks and play a number of different roles in a hypothesis recognition phase.
