Separating Pistachio Varieties Using Automatic Trainable Classifier

In this study an automatic trainable classifier, based on artificial neural network (ANN), for separating four different varieties of pistachio nuts is presented. In total 3200 sample data were selected in the development of the ANN models. Signal analysis procedures from time-domain and frequency-domain as well as statistical procedures were used for feature extraction. By performing principal component analysis on the input data set, more than 98% reduction in the dimension of the feature vector was achieved. Altogether 40 features were selected as input vector to ANN models. In order to find optimal configuration, several architectures, each having different number of neurons in the hidden layer, were designed and evaluated. The ANN models were trained using the gradient descent with momentum learning rule. Optimal configuration had a 40-12-4 structure, i.e., a network having one hidden layer with 12 neurons. This selection was based on the minimization of mean square error (MSE) and correct separation rates (CSR). The estimated MSE and CSR were 0.018 and 97.5%, respectively. I.e., only 2.5 % of pistachio nuts were misclassified.