Abstract: To establish a grading method for green Arabica coffee bean based on the lightweight YOLOv8-FasterBlock model, a total of 500 g of first-grade, second-grade, third-grade and defective green Arabica coffee bean from Yunnan were collected and mixed for acquirement of RGB images as the dataset for coffee bean grading. The YOLOv8n model was improved by replacing the BottleneckBlock in the C2f module with the FasterBlock module in FasterNet, resulting in a new lightweight YOLOv8-FasterBlock model. The improved model took 2.4 ms, on average, to discriminate different grades of coffee beans with accuracy, recall, and average precision of 98.4%, 94.3%, and 97.4%, respectively. Comparison, ablation, lightweighting, and adherent bean tests proved the superiority and structural validity of the YOLOv8-FasterBlock model. The YOLOv8-FasterBlock model improved the feature extraction capacity and inference speed for green Arabica coffee bean while having reduced complexity, enabling rapid grading of coffee bean. The results of the study provide a reference for the deployment of vision module in green Arabica coffee bean grading equipment, and also provide theoretical support for the grading of other agricultural products.

Key words: Arabica coffee; green coffee beans; YOLOv8-FasterBlock model; target detection; grading