Robust Quality Clustering of Rice Varieties Using Object Detection

The classification of rice according to SNI 6128:2020 is essential for ensuring eligibility and standardizing rice quality. An automated classification system is necessary to achieve accurate, objective, fast, cost-effective, and efficient results. This study develops an automated classification system to address the diverse rice varieties in Indonesia, generally categorized as short, medium, and long. The research findings are crucial for aiding farmers in enhancing their rice production standards. The study utilized actual rice samples available to farmers, represented by a specific variety of rice from Indonesia. The research used datasets comprising short and long-grain rice as per SNI standards, while medium-grain data were sourced from available online datasets. The datasets were trained using one-stage object detection algorithms: YOLOv5-x, YOLOv7-x, and YOLOv8-x. The findings indicate that YOLOv7-x exhibits the best performance, with the short-grain model achieving a mAP0.5 of 99.39% and a mAP0.5:0.95 of 88.61%. Furthermore, the short-grain dataset demonstrates the highest accuracy, followed by medium and long-grain datasets, meanwhile, YOLOv8-x has a faster training time, 1.004 to 2.35 or 0.437% to 57.5% faster. Notably, the accuracy for the medium-grain dataset also shows improvement compared to Karlwillem’s rice quality detection using YOLOv4, where mAP50 was 98.2%.