GAN-Augmented Deep Learning Model for Automated Fruit Ripeness Classification

Abstract

Manual fruit harvesting is costly, labor-intensive, and difficult to scale. We present a vision-based system for automated fruit-ripeness classification that jointly predicts fruit type and ripeness stage as an eight-class task {mango, strawberry, tomato, sweet pepper}×{ripe, unripe}. To mitigate limited and imbalanced field data, we train class-conditioned GANs to synthesize additional images and build an augmented training set (3,988 real images → 57,339 total). Synthetic samples are quality-controlled using automated filters and human review and are used only in training; validation and test splits contain real images exclusively. We benchmark three classifiers, DenseNet-201, ResNet-50, and a compact CNN, using a unified pipeline with standard preprocessing and on-the-fly augmentation. DenseNet-201 achieves the best generalization, reaching 99.41% training accuracy, 98.01% validation accuracy, and 95.7% accuracy on the held-out real-image test set, outperforming ResNet-50 and the CNN baseline on recall, precision, and F1 score. The results indicate that targeted generative augmentation improves robustness to variations in viewpoint, partial occlusion, and illumination, enabling reliable ripeness assessment from in-orchard imagery. The proposed pipeline provides a reproducible foundation for integrating ripeness classification into robotic harvesting workflows and can be extended to additional crops, sensing modalities, and on-device inference.