A digital learning system for tracking pattern features

A novel pattern processing scheme has been investigated which makes use of the motions generated by a window which tracks the lines or contours of a digitised television image of a black/white pattern. The novel features of the proposed scheme are that adaptive learning networks are used for both tracking and classifying. The tracking strategy is learnt from a human teacher.

Here one combines two methods of machine pattern recognition which, in isolation, have a limited performance. These are, 'static learning networks' which have known limitations, and 'programmed tracking systems' in which the pre-programming itself may be limiting. In this combination one avoids some limitations of these systems because pre-programming of strategies is not necessary and feedback exists to make the task of the nets a dynamic one.

The thesis describes a hardware visual input and a special-purpose software system which were developed for this investigation. Also, several new modifications of the SLAM (Stored Logic Adaptive Microcircuit) element are discussed.

Beyond its practical application it is possible to conclude that the system developed here may be useful in the study of hypotheses regarding living animal systems which involve eye movements.