Rice classification via convolutional neural network and transfer learning

This study classifies 20 rice varieties using three custom-built CNN architectures and four transfer-learning models: MobileNetV2, VGG16, ResNet50, and DenseNet121. The Bangladesh rice dataset, comprising original and augmented sets, was used to address the challenges of limited training data and poor model generalization. The performance of the proposed custom architectures and transfer-learning models was compared using accuracy, precision, recall, and F1-score. Custom Architecture 1 achieved the highest classification accuracy of 98.51%. Among the proposed architectures, Custom Architecture 1 achieved the highest classification accuracy on the combined dataset (original and augmented training sets), compared with 87.94% on the original dataset. The results show an increase in classification accuracy for the combined dataset, although the magnitude of the change varied across architectures. ResNet50 outperformed all transfer-learning models on the original dataset, whereas VGG16 achieved the highest accuracy of 97.87% on the combined dataset. In addition, image-resolution analysis shows that increasing the dataset size together with high-resolution inputs produced the best performance when compared with other resolutions.