Analysis of visual features and classifiers for Fruit classification problem

Abstract Analysis of visual cues for fruit classification and sorting allows to automate the visual inspection and packaging process in agricultural applications that is performed so far by human workers. Challenges for automated multi-class sorting systems are similarity in color and shape of different fruit varieties and variation among the same category of fruit. A major constraint in using well known deep neural networks for fruit classification arises because deep neural networks require large training datasets for achieving high accuracies which are generally not available in case of agricultural products especially various fruits and vegetable varieties. A thorough analysis is required to find an appropriate combination of various handcrafted features that could give precise and accurate classification results for small datasets. This paper investigates the use of various handcrafted visual features for fruit classification using traditional machine learning techniques. Different color, shape and texture features are analyzed by comparing the results obtained from six supervised machine learning techniques including K nearest neighbors, Support Vector Machines, Naive Bayes, Linear Discriminant Analysis, Decision Trees and Feed forward back propagation neural network. We propose a novel combination of Hue, Color-SIFT, Discrete Wavelet Transform and Haralick features in fruit classification problem that outperforms other handcrafted visual features. This feature combination is found to be invariant to rotation and illumination effects and works well with intra class variations providing good results for identifying subcategories of fruits along with high classification accuracies obtained for difficult fruit categories that are visually similar. It is found that Color-SIFT features alone work very well for fruit classification problem by outperforming other individual handcrafted features. Our approach is trained and tested on publicly available Fruits 360 dataset. Out of different classifiers best results are obtained using Back Propagation Neural Network, SVM and KNN classifier with classification accuracies between 99% and 100%.