DHN-YOLO: A Joint Detection Algorithm for Strawberries at Different Maturity Stages and Key Harvesting Points.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Automated strawberry harvesting faces challenges like occlusion, overlap, and variable lighting affecting fruit detection and maturity assessment.
• Accurate identification of strawberry maturity and picking points is crucial for robotic harvesting efficiency.

Purpose of the Study:

• To develop an advanced computer vision model for robust strawberry detection and keypoint estimation in complex agricultural environments.
• To improve the performance of automated harvesting systems by addressing limitations in current fruit identification methods.

Main Methods:

• Constructed the MSRBerry dataset, featuring diverse strawberry varieties, ripening stages, and field conditions.
• Proposed DHN-YOLO, an enhanced YOLOv11-pose framework incorporating CDC, C3H modules, and a novel neck architecture with attention mechanisms and dual-path fusion.
• Optimized feature capture, multi-scale extraction, and perception of critical regions for improved accuracy and robustness.

Main Results:

• DHN-YOLO achieved 87.3% precision, 88% recall, and 78.6% mAP@50:95 for strawberry detection, outperforming YOLOv11-pose.
• Keypoint detection performance reached 83% precision, 87.5% recall, and 83.6% accuracy.
• The model demonstrated high inference speed (71.6 FPS) and efficiency, surpassing other mainstream models like YOLOv13, YOLOv10, and DETR.

Conclusions:

• DHN-YOLO effectively addresses the challenges of strawberry detection and keypoint identification in complex agricultural settings.
• The proposed model offers a robust and efficient solution for practical robotic harvesting applications.
• The advancements in feature extraction and fusion contribute to superior performance in variable field conditions.