A keypoint-based method for detecting weed growth points in corn field environments

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Weed growth significantly reduces corn yield, necessitating advanced weed management strategies.
• Precision weeding requires accurate detection and localization of weed growth points, especially during early growth stages.
• Field conditions like occlusion, dense growth, and variable lighting pose significant challenges for current weed detection systems.

Purpose of the Study:

• To develop a precise weed growth point detection method for corn fields.
• To enhance weed detection accuracy and robustness in complex agricultural environments.
• To create a lightweight and efficient model for real-time weed control applications.

Main Methods:

• Proposed a keypoint pose estimation model for detecting various weed species and their growth points.
• Designed a dilation-wise residual module (DWRM) to handle occlusion and dense growth.
• Incorporated a separation and enhancement attention module (SEAM) for improved pose estimation.
• Utilized the RepViT block (RVB) for model lightweighting to suit field computational constraints.

Main Results:

• The SRD-YOLO model achieved a keypoint mean average precision (mAP_kpt) of 96.5% and an F1 score of 94%.
• The system demonstrated a high processing speed with 169 frames per second (FPS).
• Model parameters were reduced by 8.7 million, indicating significant lightweighting.

Conclusions:

• SRD-YOLO effectively meets the demands for growth point localization in challenging corn field conditions.
• The developed method provides robust technical support for real-time and precise weed control in agriculture.
• This advancement contributes to more efficient and sustainable corn production through improved weed management.