A Lightweight Recognition Method for Rice Growth Period Based on Improved YOLOv5s

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning for crop monitoring

Background:

• Accurate identification of rice growth stages is crucial for optimizing yield and quality.
• Current manual observation methods are inefficient and subjective.
• Automated methods are needed to overcome the limitations of manual rice growth stage identification.

Purpose of the Study:

• To develop a lightweight and efficient deep learning model for automatic rice growth period recognition.
• To improve the speed and accuracy of rice growth stage identification.
• To reduce the computational cost and model size for practical deployment.

Main Methods:

• Proposed Small-YOLOv5, an improved YOLOv5s architecture.
• Integrated MobileNetV3 as the backbone for reduced model size and faster detection.
• Implemented GsConv, a lightweight convolution, in the feature fusion stage to decrease computational complexity while maintaining learning ability.

Main Results:

• Small-YOLOv5 achieved an 82.4% reduction in model parameters, 85.9% decrease in GFLOPS, and 86.0% size reduction compared to YOLOv5s.
• The model demonstrated a mean Average Precision (mAP) of 98.7% (at 0.5 IoU), only 0.8% lower than the original YOLOv5s.
• Compared to YOLOV5s-MobileNetV3-Small, Small-YOLOv5 showed a 10.0% parameter reduction, 9.6% volume decrease, 5.0% mAP improvement (to 94.7%), and 1.5% recall enhancement (to 98.9%).

Conclusions:

• Small-YOLOv5 offers a highly effective and superior lightweight solution for automatic rice growth period identification.
• The model's efficiency and accuracy make it suitable for practical applications in precision agriculture.
• The developed method significantly overcomes the limitations of manual observation in rice cultivation management.