Comparison of multilayer perceptron and probabilistic neural networks in artificial vision. Application to the discrimination of seeds

Abstract

In classification problems the most commonly used neural network is probably the multilayer perceptron network (MLPN). The probabilistic neural network (PNN) is a possible alternative to the MLPN. The PNN is based on the Bayesian approach and a non-parametric estimation of the probability density functions of the qualitative classes. In this paper the performances of the PNN and the MLPN were compared on an illustrative application which consisted of the discrimination of seed species by artificial vision. The colour images of individual kernels of four species (two cultivated and two adventitious ones) were acquired. A set of 73 features characterizing the seed size, shape and texture was extracted. The data collection was divided into a training set of 1600 seeds and a test set of 800 seeds. A stepwise discriminant analysis made it possible to select the first four relevant variables among the 73 available ones. The MLPN incorrectly classified 44 and 28 seeds of the training and test sets respectively. Three configurations of the PNN were tested on the same data collection. The most sophisticated version of the PNN gave 17 and 19 misclassifications in the same data sets. The PNN presents an architecture in which all the units are operating in parallel and a hardware implementation of this kind of architecture is therefore possible. All the scaling parameters of the PNN can be determined from the training set. In contrast, there is no algorithm to automatically determine the structure of the MLPN. © 1997 John Wiley & Sons, Ltd.