Classification of Pomegranate Fruit using Texture Analysis of MR Images

Abstract

Images obtained by Magnetic Resonance Imaging (MRI) of Iranian important export cultivar of pomegranate Malase-e-Torsh were analyzed by texture analysis to determine Gray Level Co-occurrence Matrix (GLCM) and Pixel Run-Length Matrix (PRLM) parameters. The T₂ slices measured at 1.5 T for 4 quality classes of pomegranate semi-ripe, ripe, over-ripe and internal defects classes were analyzed numerically using the software MaZda. To classify pomegranate into different classes, discriminant analysis was conducted using cross-validation method and texture features. Ten GLCM and 5 PRLM features were used in 2 different classifiers. Mean classification accuracy was 95.75 % and 91.28 % for GLCM and PRLM features respectively. By using GLCM and RPLM features, classification accuracy for semi-ripe, over-ripe and internal defects classes was higher when GLCM features were used. Ripe class had higher classification accuracy while PRLM features were used. To improve classification accuracy, combination of GLCM and PRLM features were used. For achieving best classification accuracy, optimum numbers of features were selected based on their contribution to the model. Combination of 7 GLCM and 4 PRLM features resulted in mean accuracy of 98.33 % and the lowest type I and II errors. Especially, type I error in ripe and over-ripe classes were significantly decreased. The classification accuracies were 100, 98.47, 100 and 95 % for semi-ripe, ripe, over-ripe and internal defects classes.