A Lightweight Deep Learning Approach for Indian Fruit Classification Using EfficientNetV2B2

This research aims to build a system that automatically categorizes good and bad fruits, for the top six popular Indian varieties of Fruits using deep learning. By adapting pre-trained deep learning models (transfer learning), We have achieved high accuracy (over 99%) in distinguishing fruit quality, with the EfficientNetv2b2 model performing best. The research demonstrates that deep learning could be used to fruit quality assessment in agriculture and the food industry. The framework is developed based on the Pooling and Convolutional layers of the MobileNetv2, ResNet50, DenseNet, and Efficient-Netv2b2 pre-trained CNNs are used for feature extraction, and newly mentioned output and connection layers are applied to classification. Test results show that EfficientNetv2b2 achieves high accuracy (training accuracy - 99.82% and test accuracy - 98.90%) on four pre-trained networks, considering sample size and testing risk. Our goal in this paper is to show how effective transfer learning is and also overcome the limitations of current Fruit Classification Models.