Using back-propagation networks to assess several image representation schemes for object recognition

Summary form only given, as follows. Two chapters of research are presented. The first constitutes a demonstration that backpropagation networks can be used as a content addressable memory for visual objects represented within digitized real-world images. For networks encoding two or three classes of traffic signs, classification generalization is demonstrated for objects at new positions on the image frame and also for new instances of a trained class of object. The new instance may even be a somewhat degraded representation. Given this optimistic introduction, the work evolves into a second, more comparative chapter. In this further probe, backpropagation networks are used as content addressable memories with which to determine the relative value of several different visual object representation schemes. These representation schemes are tested along multiple parameters to deduce the efficacy of the scheme itself, and the influence of network parameter changes on the learning and categorization of objects.