Lightweight safflower cluster detection based on YOLOv5.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Effective safflower detection is vital for automated agricultural systems like navigation and harvesting.
• Current methods struggle with small safflower size, dense distribution, and complex field conditions, leading to low accuracy and high computational cost.

Purpose of the Study:

• To develop an improved safflower target detection model, Safflower-YOLO (SF-YOLO), enhancing accuracy and efficiency for agricultural applications.

Main Methods:

• Implemented Ghost_conv and CBAM attention mechanism in the backbone network for improved efficiency and feature extraction.
• Introduced a combined loss function and K-means clustered anchor boxes for better multi-scale adaptation and faster convergence.
• Applied data augmentation (Gaussian blur, noise, sharpening, channel shuffling) for robustness.

Main Results:

• SF-YOLO reduced GFlops by 16.6% (13.2 G) and Params by 23.9% (5.34 M) compared to YOLOv5s.
• Achieved a 1.3% increase in mAP0.5, reaching 95.3% accuracy.
• Demonstrated superior performance in complex agricultural environments.

Conclusions:

• SF-YOLO significantly improves safflower detection accuracy and computational efficiency.
• The model provides a valuable reference for developing autonomous visual navigation and non-destructive harvesting technologies in safflower cultivation.