Automated Fruit Ripeness Classification in Robotic Harvesting Using a GAN-Augmented Deep Learning Framework

Ayesha Hakim

Keywords: Transfer Learning, Synthetic Image Generation, Agricultural Automation, Precision Agriculture, Computer Vision

TL;DR: Automated Fruit Ripeness Classification in Robotic Harvesting Using a GAN-Augmented Deep Learning Framework

Abstract: Manual fruit harvesting is labor-intensive and inconsistent, limiting efficiency in large-scale agricultural operations. This study proposes a Generative Adversarial Networks (GAN)-augmented dataset integrated with a deep learning framework for automated fruit ripeness classification. Four commercially important fruits, mango, strawberry, tomato, and sweet pepper, were classified into ripe and unripe categories. As collecting large numbers of labeled fruit images is difficult due to seasonal, environmental, and labor constraints, GANs were emploed to expand the dataset by creating realistic fruit images under diverse conditions. DenseNet201, ResNet50, and a CNN were trained and evaluated, with DenseNet201 achieving the best performance (99.41% training, 98.01% validation, and 95.7% testing accuracy). GAN-based augmentation improved generalization under occlusion and illumination variability. The results confirm that DenseNet201 trained on a GAN-augmented dataset provides a reliable framework for automated ripeness classification, supporting the development of robotic harvesting systems.

Submission Number: 8