Performance comparison of second order statistics texture features for variety identification of paddy seeds

Abstract

The aim of this research was to evaluate the second order statistics texture features - grey level coocurrence matrix and grey level run length using the neural network architectures for cereal grain classification. For the purpose of classifying four paddy (rice) grains, viz. Karjat-6, Karjat-2, Ratnagiri-4 and Ratnagiri-24, an evaluation of the classification accuracy of texture features and neural network was done. To extract the features from the high-resolution images of kernels of four grain types, algorithms were written and these were used as input features for classification. Different feature models were observed for their capability to categorise these cereal grains. For the accurate classification, effect of using different features was studied and the most suitable feature from the feature set was also identified. The Texture-GLCM feature set outperformed the Texture-GLRL feature set in most of the instances of classification. Also the performance of three training functions viz. Levenberg-Marquardt (LM) backpropagation, resilient backpropogation (RP) and scaled conjugate gradient (SCG) training functions was compared and the most reliable training functions was identified from the three functions for accurate classification of four paddy varieties.