R. L. Cromwell, A. C. Kak
Object recognition requires complicated domain-specific rules. For many problem domains, it is impractical for a programmer to generate these rules. A method for automatically generating the required object class descriptions is needed - this paper presents a method to accomplish this goal. In our approach, the supervisor provides a series of example scene descriptions to the system, with accompanying object class assignments. Generalization rules then produce object class descriptions. These rules manipulate non-symbolic descriptors in a symbolic framework; the resulting class descriptions are useful both for object recognition and for providing clear explanations of the decision process. We present a simple method for maintaining an optimal description set as new examples (possibly of previously unseen classes) become available, providing needed updates to the description set. Finally, the system’s performance is shown as it learns object class descriptions from realistic scenes - video images of electronic components.