Cgc-YOLO: A New Detection Model for Defect Detection of Tea Tree Seeds

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Tea tree seeds are highly sensitive to dehydration, necessitating effective methods for quality assessment.
• Current methods for detecting seed defects are insufficient for small-scale and complex surface imperfections.

Purpose of the Study:

• To develop an enhanced YOLO-based model, Cgc-YOLO, for accurate and efficient detection of surface defects in tea seeds.
• To improve the preservation of tea seed germination rates and overall quality through advanced defect detection.

Main Methods:

• A high-resolution imaging system was used to create a dataset of five tea tree seed types with diverse defects.
• Cgc-YOLO integrates GhostBlock in the Backbone for efficient feature extraction and CPCA attention in the Neck for enhanced detail sensitivity.
• The model was trained and evaluated on the custom tea seed defect dataset.

Main Results:

• Cgc-YOLO achieved high performance metrics: 97.6% mAP50 and 94.9% mAP50-95.
• The model demonstrated superior accuracy compared to YOLO11, with improvements of 2.3% and 3.1% in mAP50 and mAP50-95, respectively.
• Cgc-YOLO maintains a compact model size of 8.5 MB, balancing accuracy with efficiency.

Conclusions:

• Cgc-YOLO offers a robust and lightweight solution for the nondestructive detection of tea seed surface defects.
• The proposed model contributes to intelligent seed screening and enhances quality assurance in seed storage.
• This research provides a valuable tool for improving the viability and marketability of tea seeds.