Rice Quality Identification Based on Transfer Learning

Rice is crucial for global food security and is among the most vital crops worldwide. Given its diverse types and characteristics, accurately classifying rice is essential for genetic improvement and rice breeding. Traditional methods and machine learning techniques for rice classification often fall short in terms of accuracy and efficiency. Hence, this paper proposes a rice type classification method based on convolutional neural network and transfer learning. The study utilizes the Rice Image Dataset, comprising 15,000 images per type across 5 rice varieties. Transfer learning is employed to reduce training costs, so we utilized pretrained models such as AlexNet, VGG16, ResNet18, MobileNetV2, and DenseNet121 to classify the dataset. Experimental results indicate that all five CNN models effectively classify rice types. However, MobileNetV2 stands out with the highest accuracy, F1-score, and kappa coefficient, along with the lowest parameters and floating-point operations (FLOPs), demonstrating superior accuracy and efficiency. These findings hold significant implications for selecting appropriate models in real-world agricultural applications, aiding in more precise rice classification.