Comparative analysis of YOLO models for green coffee bean detection and defect classification

Area of Science:

• Computer Vision
• Agricultural Technology
• Food Science

Background:

• Green coffee bean quality directly impacts flavor and market value.
• Automated identification of bean types and defects offers significant industry advantages.
• Accurate defect detection is crucial for maintaining product consistency and consumer satisfaction.

Purpose of the Study:

• To evaluate the performance of various You Only Look Once (YOLO) models for green coffee bean identification and classification.
• To compare different YOLO variants based on computational efficiency, accuracy, and speed.
• To identify the most effective YOLO model for real-world implementation in coffee bean quality control.

Main Methods:

• A dataset of 4,032 training and 506 testing images featuring diverse green coffee beans, defects, and lighting conditions was utilized.
• Multiple YOLO variants (YOLOv3, YOLOv4, YOLOv5, YOLOv7, YOLOv8) and custom models were trained and evaluated.
• Performance metrics including precision, recall, f1-score, and mean average precision (mAP) were assessed.

Main Results:

• The custom-YOLOv8n model demonstrated superior performance in accuracy, precision, recall, and mAP.
• Bounding boxes accurately encompassed coffee beans, with uniform black backgrounds facilitating detection.
• The fine-tuned custom model effectively distinguished bean types and detected subtle defects.

Conclusions:

• The custom-YOLOv8n model is highly effective for automated green coffee bean quality control.
• This research provides valuable insights for selecting and implementing YOLO models in agricultural systems.
• Optimized YOLO models can significantly enhance efficiency and consistency in the coffee industry.