Banana cultivar classification using scale invariant shape analysis

Abstract

This paper presents scale-invariant shape analysis with respect to banana cultivar detection. We consider three cultivars: Cavendish, Lady Finger, and Pisang Awak. We present the appropriate image preprocessing methods and compare different feature selection algorithms, numbers of features, as well as various machine learning models as the classifier. We found that the best feature selection method is chi square, and the optimal number of features is 100. Differences between prediction accuracy of machine learning models are small, but overall, Bayesian network performs the best, with overall AUC of 0.933 and overall accuracy of 84%.