RP-DETR: end-to-end rice pests detection using a transformer

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Pest infestations in rice crops significantly impact yield and quality, necessitating early detection.
• Visual inspection is crucial, and precision agriculture demands fast, accurate automatic pest identification.
• Deep learning models are favored for plant pest detection due to their high efficiency.

Purpose of the Study:

• To develop an effective rice pest detection framework using the Transformer architecture.
• To enhance feature extraction efficiency and reduce model parameters.
• To improve multi-scale feature fusion and detection performance.

Main Methods:

• Utilized Transformer architecture for capturing long-range features.
• Introduced RepPConv-block to reduce information redundancy and model parameters.
• Integrated Gold-YOLO neck into the CCFM structure for improved multi-scale feature fusion.
• Employed an MPDIoU-based loss function to enhance detection performance.

Main Results:

• The proposed RP18-DETR and RP34-DETR models reduced parameters by 16.5% and 25.8% respectively.
• Achieved higher identification accuracy on a self-constructed rice pest dataset.
• RP18-DETR showed a 1.2% higher average accuracy compared to the original RT18-DETR at a 0.5 threshold.

Conclusions:

• The developed framework effectively detects rice pests with improved accuracy and reduced model complexity.
• The enhancements contribute to more efficient and reliable pest identification in precision agriculture.
• This approach offers a promising solution for mitigating yield losses due to pest infestations.