FNE-RTDETR: A lightweight end-to-end model for small-area tomato leaf disease detection and fine-grained...

Area of Science:

• Plant Pathology
• Computer Vision
• Agricultural Technology

Background:

• Tomato cultivation is economically significant but vulnerable to environmental factors impacting yield and quality.
• Accurate and timely detection of tomato leaf diseases is crucial for maintaining crop production.
• Current disease detection methods are subjective, inefficient, and struggle with the balance between model lightness and accuracy.

Purpose of the Study:

• To introduce the FNE-RTDETR model for efficient and accurate tomato leaf disease detection.
• To address the challenge of achieving high recognition accuracy with lightweight models in natural environments.
• To improve feature extraction efficiency and fine-grained classification of various disease types.

Main Methods:

• Developed the FNE-RTDETR model by replacing the RT-DETR backbone with FasterNet and integrating lightweight PConv.
• Incorporated Deformable attention and AIFI modules to enhance fine-grained classification.
• Replaced the decoder's cross-attention mechanism with an efficient multi-channel attention mechanism for improved multi-scale feature fusion.

Main Results:

• FNE-RTDETR achieved an mAP50 of 91.5%, outperforming RT-DETR by 4.1%.
• The model reduced parameters by 14.7% and GFLOPs by 9% compared to RT-DETR.
• Demonstrated superior convergence speed, generalization ability, and lightweight performance across multiple datasets.

Conclusions:

• The FNE-RTDETR model offers a significant advancement in tomato leaf disease detection.
• It effectively balances lightweight design with high detection accuracy, outperforming existing models like YOLO variants and DETR.
• The model shows strong potential for practical application in precision agriculture for timely disease management.