Focus of attention in image understanding systems

An image understanding system (IUS) was developed for the processing of aerial photographs into cartographic images. The system makes use of both a bottom-up (data driven) and a top-down (object driven) analysis scheme. The bottom-up analysis makes use of a classical pattern recognition approach. This is done by implementing separate object recognition algorithms to obtain regions that will serve as the foci of attention (FOA). The modular composition of the system makes the extensibility of the software a major feature of the system. A scene description obtained from the bottom-up analysis is then analyzed by the top-down analysis engine. This scheme consists of a rule base, which describes the relationship between the objects in the scene. The top-down analysis is very important in that it reduces the false alarm rate (FAR) of the system. The IUS was implemented using a combination of serial computers and transputer networks. The transputer network implemented all the low-level image processing functions, while the serial computers were left free to selectively apply the rules to the various reported objects. The IUS developed is demonstrated on a number of synthetic images to show the FAR in the case of an exactly determinable image. A true aerial photograph is also analyzed and the quality of the analysis is discussed in relation to existing systems