Design of field real-time target spraying system based on improved YOLOv5

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Deep learning excels at identifying plants in complex agricultural settings.
• Precision spraying systems are crucial for efficient and targeted pesticide application.

Purpose of the Study:

• To design and develop a machine vision precision real-time targeting spraying system for field scenarios.
• To improve the efficiency and accuracy of weed detection and spraying using deep learning.

Main Methods:

• Proposed a system architecture including image acquisition, recognition, electronic spraying, and pesticide supply modules.
• Improved the YOLOv5s deep learning model by modifying the backbone network and incorporating an attention mechanism.
• Developed a grille decision control algorithm for solenoid valve operation and trained the model on a dataset of common weeds.

Main Results:

• The improved algorithm reduced model size by 46.43% while increasing FPS by 18.16% with minimal impact on mAP accuracy.
• Achieved on-target spraying accuracies of 90.80% at 2km/h, 86.20% at 3km/h, and 79.61% at 4km/h.
• Found that spraying accuracy decreased with increased operating speed, primarily due to reduced effective recognition rates.

Conclusions:

• The developed system effectively integrates deep learning for real-time precision spraying in agriculture.
• The optimized deep learning model offers a balance between reduced size, increased speed, and maintained accuracy.
• Operating speed significantly influences spraying accuracy, highlighting the need for adaptive control strategies.