A study of morphological feature detector complexity and character recognition rates

A structural complexity measure that is useful for generating morphological feature detectors is described. The question of how to assess a complexity measure is addressed. The approach is to define a specific complexity measure and to investigate its correlation with performance measures. Factoring this type of information into search strategies offers the promise of more efficient algorithms for designing structuring elements. Two other basic questions are addressed: the optimal performance levels for single detectors; and the problem of generalising the performance when a detector is confronted with new samples of handwritten letters. The complexity measure is evaluated using two-class handwritten character recognition experiments. Results suggest that there is a complexity band that can be used to aid in the search for generalizable feature detectors