Abstract:
This thesis presents a deep learning–based approach for automating the classification of
fruits as fresh or rotten, focusing on apples, bananas, and oranges. A dataset of over 13,000
images was preprocessed through resizing, normalization, and augmentation before being
used to train and evaluate multiple architectures, including VGG16, ResNet50,
InceptionV3, MobileNetV2, Xception, and a hybrid model. The results revealed that the
Trained Hybrid Model and InceptionV3 outperformed others, with validation accuracies of
96.52% and 91.89%, respectively, demonstrating strong generalization ability. Results
showed reliable detection of rotten fruits but weaker performance for fresh fruits,
highlighting the importance of balanced datasets and improved optimization. The study
demonstrates the potential of deep learning for smart agriculture, offering insights into
model performance and practical deployment, and suggests future improvements such as
larger datasets, advanced architectures, mobile integration, and explainable AI for
trustworthy decision making.
