Plant growth point localization via epoch-based prior annealing

Area of Science:

• Agricultural Engineering
• Computer Vision
• Plant Science

Background:

• Precision agriculture requires accurate localization of plant growth points for tasks like automated weeding.
• Current methods often focus on object detection, lacking fine-grained growth point identification.

Purpose of the Study:

• To develop a novel training strategy for improved keypoint regression of plant growth points.
• To enhance the accuracy and robustness of plant detection models for precision agriculture.

Main Methods:

• Proposed epoch-based prior annealing (EPA) strategy incorporating the excess green minus excess red (ExG-ExR) index.
• Introduced schedule and gain factors to guide keypoint regression.
• Evaluated performance using mean average precision (mAP) metrics.

Main Results:

• EPA improved keypoint localization performance, increasing mAP50 by 0.024 and mAP50:95 by 0.011.
• Bounding box detection performance was maintained.
• Additional computational cost of EPA was negligible (<0.5%).

Conclusions:

• The EPA strategy significantly enhances accuracy, robustness, and generalizability in plant and growth point detection.
• EPA offers a practical and scalable solution for precision agriculture applications.
• Optimal guidance strength is crucial for effective training with EPA.