A decomposition machine-learning strategy for automated fruit grading

Automated grading of fruit is an important industrial task that is expanding rapidly in its uptake. Machine learning-based techniques are increasingly being applied to this domain in order to formulate effective solutions for complex classification tasks. The inherent variability in the visual appearance of fruit and its quality-determining features, contributes to it often being a challenging classification task with much potential for improving the predictive accuracies for many fruit varieties. Additionally, the usability of many sophisticated machine learning algorithms in the form of tunable parameters and interpretable outputs is low, thus presenting a real barrier for the uninitiated. We address these problems by decomposing the overall machine learning task into subproblems. We propose combining a more sophisticated boosting algorithm (AdaBoost. ECC) with low interpretability for the learning of fruit-surface characteristics, whose outputs can then be combined with rule induction algorithms (RIPPER and FURIA) that learns the overall fruit grading rules with outputs of high interpretability for the operators to both review and revise. Our initial experiments considered four fruit datasets. We compared the results of our approach with that from a commercial system using manual calibration of the fruit grading parameters and found that our strategy can improve the accuracy over the current industry methods while providing high usability and interpretability of outputs.