<![CDATA[The manual assessment of banana ripeness on an industrial scale is subjective, time-consuming, and inconsistent. This necessitates an automated computer vision system. Previous studies have used shallow Convolutional Neural Networks (CNNs) for binary classification, but these networks often struggle with complex ripening stages and degrade in deeper networks. This study addresses this gap by implementing a deep learning algorithm using the ResNet-50 architecture. The residual block mechanism extracts fine-grained visual features without vanishing gradient issues. The model was evaluated using a diverse dataset of 13,478 digital images spanning four stages of banana ripeness: overripe, ripe, rotten, and unripe. Using a 95-5 train and test-validation split, the model was optimized over 50 epochs with a categorical cross-entropy loss function. The proposed model achieved outstanding accuracy (98.13%), minimal loss (0.1237), and average precision, recall, and F1-scores of 98.09%, 98.20%, and 98.14%, respectively. This study scientifically validates the robustness of deep residual networks in complex agro-industrial pattern recognition. Furthermore, with an inference time of approximately 50 ms per image, the system is ready for seamless integration into an automated sorting line.]]>
Copyrights © 2026
