Precision citrus segmentation and stem picking point localization using improved YOLOv8n-seg algorithm.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic harvesting of citrus is challenging due to small stem size, background color similarity, and variable fruit positioning.
• Accurate localization of picking points is crucial for efficient and effective automated citrus fruit harvesting.

Purpose of the Study:

• To develop an improved YOLOv8n-seg model for segmenting citrus fruits and stems.
• To achieve accurate localization of picking points for robotic harvesting using geometric constraints.
• To enhance feature representation and small-object detection for improved segmentation accuracy.

Main Methods:

• Replaced standard convolutions with GhostConv to reduce model parameters.
• Integrated a convolutional block attention module (CBAM) and a small-object detection layer.
• Incorporated fruit-stem positional relationships and geometric constraints for stem matching and optimal picking point determination.

Main Results:

• Achieved recall rates of 90.91% for fruits and stems, with precision rates of 96.04% (fruit) and 97.12% (stem).
• Obtained a mean average precision (mAP50) of 94.43% and an F1-score of 93.51%.
• Demonstrated high real-time performance with an average detection rate of 88.38% for picking points within 373.25 milliseconds (GPU supported).

Conclusions:

• The improved YOLOv8n-seg model reliably and effectively localizes citrus picking points.
• The study provides a strong technical foundation for the advancement of automated citrus fruit harvesting systems.