A Hybrid Human-in-the-Loop Framework for Industrial Cocoa Bean Grading Using YOLOv11 and Vision Transformers

Manual cocoa bean inspection is subjective and prone to error, significantly impacting chocolate production. This study presents an automated, high-throughput cocoa bean grading system leveraging YOLOv10, YOLOv11, and Vision Transformers (ViT). Evaluated on a diverse dataset of Thai cocoa beans (8,033 high-resolution images), the YOLOv10b model achieves state-of-the-art color classification with a mean Average Precision at 0.5 Intersection over Union (mAP50) of 0.896. For defect detection, YOLOv11x reached 0.567 mAP50. For the final deployment, the system utilizes a YOLOv11 and Vision Transformer (ViT-B/16) hybrid model, which substantially improves accuracy by 16.35% for color classification and 17.4% for defect classification. We integrated a web-based Human-in-the-Loop (HITL) platform utilizing Active Learning to iteratively refine dataset quality. Crucially, the automated system explicitly separates latency into three distinct metrics: 1) AI inference time (28.2 seconds), 2) mechanical sorting cycle time (4.53 minutes), and 3) full end-to-end evaluation time. Overall, the system reduces the full end-to-end evaluation time by 83% (from 30 minutes to 5 minutes per 100-bean sample), while establishing a highly standardized and reproducible evaluation process. This end-to-end framework provides a scalable, industrially viable solution that standardizes agricultural quality control and reduces operational costs for processing facilities.