A neural network for invariant object recognition in a roboticenvironment

摘要

Summary form only given, as follows. Object recognition, which maybe subject to occlusion or to various combinations of scaling,translational, and rotational transformations from prestored objectmodels, is under investigation. Such an environment is very typical inthe applications of robotics. A `pure' neural network approach isadopted here, i.e. without including any mathematical transforms, suchas polar or Fourier transforms, as a preprocessor. Detailed discussionson the neocognitron by Fukushima are given which show that the networkmodel is able to solve the problems of invariant recognition and ofocclusion resolving by adjusting the parameters of both staticstructures and dynamic learning rules