Shape recognition combining mathematical morphology and neural networks

Abstract The purpose of this paper is to show that son adapted preprocessing can help neural networks in classifying binary planar shapes. After recalling some basic notions of mathematical morphology and some shape numbers,a morphological preprocessing transforming a binary planar shape into a curve which is translation, rotation andscale invariant, namely the spectral function, is presented. Then, these extracted shape features are discriminatedby statistical and neural methods. Finally an application on real data is described.Key words : Mathematical morphology, spectral function, neural network. 1. Introduction In the field of shape discrimination and recognition, the direct use of neural networks is sometimes tedious due tothe huge amount of unstructured data. A multitude of techniques has been developed in the recent years for shapedescription, but we show that mathematical morphology [10, 16] brings some very good solutions to the problemof shape discrimination.The principle, [1O} is to synthetise the information contained in a shape into a curve which is translation, rotationand scale invariant, and which gives a global information of the shape. (It uses mathematical morphological idea