Convolutional Neural Networks Performance Investigation in Banana Ripeness Classification: Impact of Model, Padding, and Optimizer

Manual identification of banana ripeness often relies on subjective human perception, which can result in quality variations and inconsistent sorting, especially on a large scale. This research aims to investigate the application of automatic technology using a simple Convolutional Neural Network (CNN) in banana ripeness classification. The contributions of this research include investigating the effects of the number of hidden convolutional layers, padding type, and optimizer on the performance of a simple CNN in identifying four levels of banana maturity - namely unripe, half-ripe, ripe, and overripe - as well as comparing the performance of this simple CNN with that of a Densely Connected Convolutional Network with 201 layers (DenseNet201) and the Visual Geometry Group Network 16 (VGG16). This research proposes augmentation techniques on the training set and validation set, as well as four test scenarios involving various combinations of simple CNN configurations. The DenseNet 201 and VGG16 models were also compared to assess their performance against simple CNN. The research results show that the model with a combination of three convolutional layers, 'same' padding, and the Root Mean Square Propagation (RMSprop) optimizer gives the best results across all metrics - accuracy, precision, recall, and F1-score - with each achieving 99%. Meanwhile, when all models use two convolutional layers, the simple CNN outperforms both DenseNet201 and VGG16 across all evaluation metrics in the testing set.