Object Recognition in Image Sequences and Robust Associative Image Memory using the Multilevel Hypermap Architecture

The introduced system for object recognition and tracking uses an associative memory for storing proto-types of objects. The Multilevel Hypermap Architecture (MHA), a self-organizing neural network approach, is used, to construct a robust system. To process form vari-ant objects the MHA is extended to work with masked input data. Because of using scaled input objects, the system is in-variant to translation. The invariance to rotation is real-ized by the associative memory, which is able to learn different instances of the same input object. In our tests we obtained a robust system behavior, be-cause the associative memory is able to minimize distur-bances in feature extraction with the learned and recalled features of an object prototype