BR-YOLOv9: a multi-scale fusion framework for robust weed recognition in rice paddy fields

Area of Science:

• Agricultural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Intelligent weed recognition faces challenges in paddy fields due to varying illumination, occlusion, and complex backgrounds.
• Existing models exhibit low accuracy and limited generalization under these conditions.

Purpose of the Study:

• To develop an improved weed detection model for complex paddy-field environments.
• To enhance the accuracy and generalization capabilities of weed recognition systems.

Main Methods:

• An improved YOLOv9 model, BR-YOLOv9, was proposed.
• Incorporated a weighted bidirectional feature pyramid network (BiFPN) for enhanced feature fusion.
• Introduced a novel RepFEL module with feature enhancement and a large selective kernel mechanism.

Main Results:

• BR-YOLOv9 achieved mAP@0.5 of 98.21% and mAP@0.5:0.95 of 95.97%, outperforming the original YOLOv9.
• Demonstrated stable detection performance across diverse conditions: varying distances, illumination, and target densities.
• Exhibited superior robustness and generalization in complex agricultural settings.

Conclusions:

• BR-YOLOv9 offers a robust visual perception solution for paddy-field weed detection.
• The model provides technical support for precision weed management and intelligent weeding equipment.