Coffee Defects Detection Based on Green Bean Images Using YOLO Architecture

The quality of coffee beans significantly impacts the overall value of coffee production, and detecting defective beans is a crucial step in ensuring high standards. The traditional coffee bean quality assessment process is time-consuming and subjective, relying heavily on the expertise of a Q grader. However, there is an alternative approach using deep learning technology to develop more portable and trainable tools that can achieve high accuracy. This study compares the performance of two deep learning-based object detection methods, YOLOv8 and YOLOv10, for detecting defects in green Arabica coffee beans, categorized into two different types of primary defects. The research methodology involved collecting data, preprocessing, designing models, training, and testing. The dataset consists of 700 images of coffee beans divided into two classes, brown beans and dried cherries. YOLOv8n achieved a test accuracy of mAP50 0.995 and mAP50-95 0.875, while YOLOv10n attained a test accuracy of mAP50 0.992 and mAP50-95 0.859. In terms of computational efficiency, YOLOv8n completed training in 0.499 hours, whereas YOLOv10n finished in 0.479 hours. Despite the marginal difference in accuracy, both models demonstrate high potential for real-time detection in the coffee industry, with YOLOv8n slightly outperforming YOLOv10n in accuracy, but YOLOv10n being slightly faster. These results suggest that both methods can be effectively utilized for defect detection, offering scalable solutions for the coffee processing industry, particularly in the sorting process, to ensure the production of high-quality, valuable coffee beans.