RiceKernelEngine: Benchmarking transfer learning models for microscopic images of rice kernel

Abstract

Rice serves as a principal dietary staple food nationwide. The high demand for this cereal grain has led to extensive research, resulting in the frequent development of new grain varieties. Proper cultivation time, region, and nurturing techniques specific to each rice variety can significantly boost production. However, the subtle differences distinguishing one rice grain from another make the classification of rice kernels a resource-intensive and exhaustive task. Our proposed approach explores the potential for automated classification of five different rice varieties developed in Bangladesh: BINADHAN-8, BINADHAN-23, BRRI-67, BRRI-74, and BRRI-102. These varieties are highly similar in appearance, and state-of-the-art transfer learning models have been employed to assess their practical feasibility. The dataset comprises 3155 images, with classes containing 605, 578, 642, 695, and 635 images, respectively. The implemented models—VGG19, MobileNetV3, EfficientNet, and ConvNeXt—achieved accuracies of 85%, 94%, 96%, and 87%, respectively, with an average accuracy of 90.50%. The results indicate the practical applicability of these models in this field, with EfficientNet providing the highest accuracy rate for classifying rice grains. This study utilized an exclusive, self-obtained dataset of microscopic rice grain images. An autonomous and efficient classification system will benefit rice farmers and open new research opportunities for agriculturists and other stakeholders.