Detecting Coffee Bean Types using YOLOv11 Architecture

Abstract

Indonesia ranks fourth as the world's largest coffee producer with production reaching 789,000 tons annually, dominated by Arabica and Robusta varieties. The classification of these two varieties at the industrial level currently relies on highly subjective manual visual inspection. Conventional CNN methods are unable to provide object coordinate localization information essential for industrial actuator systems to perform automated physical separation on high-speed production lines. Thus, this study implements the YOLOv11m architecture to detect and classify coffee bean types using 3,705 images. The model demonstrates highly precise detection performance, achieving a Mean Average Precision ([email protected]) of 98.5%, F1-Score of 97.8%, precision of 97.9%, and recall of 97.8%, with an inference speed of 19.6 milliseconds per image, enabling processing of more than 38 frames per second. The YOLOv11m-based classification system delivers an accurate and efficient detection solution suitable for direct integration into automated conveyor sorting machines within the smart agriculture industry.