PHRF-RTDETR: a lightweight weed detection method for upland rice based on RT-DETR

Area of Science:

• Agricultural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Weeds in upland rice fields significantly reduce yield and quality.
• Current weed detection methods lack a balance between accuracy and lightweight design, hindering practical application.
• Intelligent weeding technologies require efficient and accurate weed detection systems for upland rice.

Purpose of the Study:

• To develop a lightweight and accurate weed detection model for upland rice.
• To enhance the RT-DETR model for improved performance in agricultural scenarios.
• To address the limitations of existing weed detection technologies in upland rice cultivation.

Main Methods:

• Proposed a novel lightweight backbone network (PGRNet) to replace RT-DETR's original feature extraction.
• Integrated HiLo mechanism into the AIFI module for enhanced multi-frequency feature capture.
• Optimized the RepC3 block to RetC3 for balanced feature fusion and computational efficiency, and replaced GIoU loss with Focaler-WIoUv3 loss.

Main Results:

• PHRF-RTDETR achieved high performance metrics (92% precision, 85.6% recall, 88.2% mAP50) with minimal deviation from the baseline.
• Demonstrated significant reductions in computational resources: 59.3% fewer FLOPs, 53.7% fewer parameters, and 53.9% smaller model size.
• Outperformed traditional models (Faster R-CNN, SSD, YOLO, RT-DETR) in balancing lightweight design and accuracy for upland rice weed detection.

Conclusions:

• The PHRF-RTDETR model offers a viable solution for weed detection in intelligent weeding robots for upland rice.
• Potential to reduce agricultural production costs through enhanced labor efficiency.
• Contributes to improved food security in drought-prone regions by optimizing rice cultivation.