Towards Flexible Recognition via Structural Description

The purpose of this report is to discuss on a possible approach towards flexible object recognition. The rationale for this approach may be best shown by an example. Suppose that we want to recognize a pair of scissors. There should be hundreds of different shapes of a pair of scissors, so we can not do it just using a template matching. We have to describe the given figure as, for instance, “two elongated elements crossed each other at the middle portionof them; one end of each element is a sharp tip and at the other end it has a loop.” Then we can recognize the figure as a pair of scissors by consultinga database of such descriptions. Thus we have to extract and identify those components such as “an elongated element” or “a loop” and their spatial relations from the given figure. We will call those components “meaningful parts.” Thus in our scheme the process should be in three stages: