Rice spikelet flowering-state detection based on an improved YOLOv11n model

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Accurate identification of rice spikelet flowering status is vital for hybrid rice breeding programs.
• Automated detection is challenging due to small target size, dense distribution, and background interference.

Purpose of the Study:

• To develop an improved AI model for accurate rice spikelet flowering status detection.
• To address the challenges of small targets, dense distribution, and background interference in automated detection.

Main Methods:

• Proposed YOLO11n-ACDW, an enhanced YOLOv11-based model.
• Incorporated ADown downsampling module for high-frequency feature preservation.
• Introduced Convolution-Attention Fusion Module (CAFM) for improved target-background discrimination.
• Utilized Detect_Efficient detection head and Wise-IoU v3 loss function for scale adaptation and regression accuracy.

Main Results:

• Achieved 88.8% precision, 84.7% recall, and 84.0% mAP@0.5 on the test set.
• Outperformed baseline YOLOv11 by 8.0% in precision, 3.4% in recall, and 1.3% in mAP@0.5.
• Surpassed RT-DETR by 9.2% in precision, 6.1% in recall, and 2.5% in mAP@0.5.
• Maintained lightweight characteristics with 2.14 M parameters and 4.4 GFLOPs.

Conclusions:

• The YOLO11n-ACDW model offers significant improvements in rice spikelet flowering status detection.
• The model provides effective technical support for flowering-window monitoring in hybrid rice breeding.
• The lightweight design makes it suitable for practical applications in agricultural settings.