A lightweight MHDI-DETR model for detecting grape leaf diseases.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate grape leaf disease diagnosis is vital for agriculture.
• Existing methods struggle to balance lightweight design with high detection accuracy.
• Need for efficient, real-time disease detection models in farming.

Purpose of the Study:

• To develop a real-time, end-to-end, lightweight grape leaf disease detection model.
• To improve upon existing detection architectures like RT-DETR for agricultural applications.
• To achieve dual optimization of model efficiency and detection accuracy.

Main Methods:

• Introduced MHDI-DETR, an improved RT-DETR architecture utilizing MobileNetv4 for backbone lightweighting.
• Developed a lightSFPN feature fusion structure integrating Hierarchical Scale Feature Pyramid Network and UniRepLKNet's Dilated Reparam Block.
• Incorporated Efficient Local Attention and Focaler-GIoU (combining GIou and Focaler-IoU) for enhanced feature capture and small target detection.

Main Results:

• MHDI-DETR achieved a 56% reduction in parameters and a 49% decrease in floating-point operations compared to RT-DETR.
• Achieved high precision rates: 96.9% for accuracy, 92.6% for mAP50, and 72.5% for mAP50:95.
• Demonstrated improvements of 1.9%, 1.2%, and 1.2% in accuracy, mAP50, and mAP50:95, respectively, over RT-DETR.

Conclusions:

• MHDI-DETR offers superior detection accuracy and a significantly lighter model compared to RT-DETR and other mainstream models.
• The model provides an efficient technical solution for automated agricultural disease management.
• Achieved dual optimization in efficiency and accuracy for grape leaf disease detection.