CNN Model with Parameter Optimisation for Fine-Grained Banana Ripening Stage Classification

Abstract

Fruit grading is a significant problem in the fruit industry because each maturity stage of the fruit represents a distinct economic worth. Banana is one of the most mass-produced fruits that must be visually classified. However, because human eye perception varies, precise classification using a machine is necessary to standardise the grading system. This research develops a four-layered CNN deep-learning model to classify bananas into seven ripening stages. To train the model, we employed Mazen and Nashat dataset and expanded it using data augmentation techniques. As a baseline, we trained a basic four-layer CNN model and achieved 88.2% of accuracy in fine-grained categorisation due to the similarity of the adjacent ripening class. To enhance the accuracy of our basic model, we applied a parameter optimisation approach to get the best hyper-parameters for the profound banana ripeness indicator. As a result, the time-constrained parameter optimisation method that we utilised successfully increased the model accuracy up to 91.2% and the F1 score at 90.5%, which is satisfactory for fine-grained banana classification compared to the previous research.