Optimization of Convolutional Neural Network Architecture for Accurate Banana Ripeness Classification

Abstract

Ripening of bananas is a crucial element that decides on the quality, taste, and prices of the fruit in the market. The traditional method for measuring ripeness is still visual inspection, which is subjective and likely to produce inconsistencies, particularly during mass sorting. The current paper focuses on developing a banana ripeness classification model using a Convolutional Neural Network (CNN) optimized with Batch Normalization, Dropout Layer, ReduceLROnPlateau, and Early Stopping. The sample dataset is made up of three types of banana images, that is, unripe, ripe, and rotten, obtained with the help of open sources and direct photography. A data augmentation process was carried out to improve the model's generalization to changes in lighting conditions and image orientation. The experimental findings show that the optimized CNN model achieved the best validation accuracy of 97.50, precision of 94.87, recall of 95.11, and F1-score of 94.99, indicating balanced performance across all classes. Early Stopping was effective in decreasing computation time (126 minutes) to 104 minutes without a significant loss in performance and avoiding overfitting. The system, as an implementation, includes an interactive Graphical User Interface (GUI) that enables users to upload images and view classification results in real time. This set of results demonstrates that the optimized CNN model can readily and quickly identify characteristic visual features of each banana ripening stage and can serve as an automated sorting system in smart agriculture.