Tutorial: Algorithms for 2-Dimensional Object Recognition.

Representation of arbitrary shape for purposes of visual recognition is an unsolved problem. The task of representation is intimately constrained by the recognition process and one cannot be solved without some solution for the other. We have already done some work on the use of an associative neural network system for hierarchal pattern recognition of the sort that may be ultimately useful for generic object recognition. The conclusions from this work were that

@ Networks can be constructed which are robust to noisy and missing data.

@ The input to the network should preferably be signi_cance measures of homogenous features.

@ The required invariance properties must be explicit in the form of input representation.

We restrict here the recognition task to pre-segmented rigid bodies in the hope that a solution for this case will suggest ways of solving the more general case.