A neural network for invariant object recognition in a robotic environment

Abstract

Summary form only given, as follows. Object recognition, which may be subject to occlusion or to various combinations of scaling, translational, and rotational transformations from prestored object models, is under investigation. Such an environment is very typical in the applications of robotics. A 'pure' neural network approach is adopted here, i.e. without including any mathematical transforms, such as polar or Fourier transforms, as a preprocessor. Detailed discussions on the neocognitron by Fukushima are given which show that the network model is able to solve the problems of invariant recognition and of occlusion resolving by adjusting the parameters of both static structures and dynamic learning rules.<>