Abstract

Banana ripeness classification is an important task in agriculture and food processing that supports automatic and objective sorting processes. Advances in deep learning, particularly Convolutional Neural Networks (CNNs), have demonstrated high capability in image classification tasks. This research aims to analyze and compare the performance of three CNN architectures, namely VGG16, VGG19, and EfficientNetB5, for image-based banana ripeness classification, considering both classification performance and computational efficiency. The dataset used was the Banana Ripeness Classification Dataset, consisting of 1,080 images categorized into three classes: unripe, ripe, and overripe. The dataset was divided into training (70%), validation (15%), and testing (15%) sets. All models were implemented using a transfer learning approach with ImageNet pretrained weights and trained for 30 epochs using the Adam optimizer. Model performance was evaluated using accuracy, precision, recall, and F1-score, while computational performance was assessed based on the total number of parameters, model size, and average inference time per image. The results show that VGG16 and VGG19 achieved the best classification performance, obtaining an accuracy of 99.38%, precision of 0.99, recall of 0.99, and F1-score of 0.99. In contrast, EfficientNetB5 achieved an accuracy of 91.36%, precision of 0.92, recall of 0.91, and F1-score of 0.91. In terms of computational performance, VGG16 contained 15.24 million parameters, with a model size of 58.2 MB and an average inference time of 32.6 ms per image. VGG19 contained 20.29 million parameters, with a model size of 77.4 MB and an average inference time of 39.8 ms per image, while EfficientNetB5 contained 30.50 million parameters, with a model size of 116.8 MB and an average inference time of 45.7 ms per image. These findings demonstrate that VGG16 and VGG19 provide superior classification performance compared with EfficientNetB5 on the utilized dataset, while VGG16 also offers the lowest computational requirements among the evaluated architectures. Therefore, VGG16 represents a promising architecture for developing an accurate and computationally efficient image-based banana ripeness classification system for automated sorting applications.