Efficient deep learning-based tomato leaf disease detection through global and local feature fusion.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Intelligent agriculture faces challenges in tomato leaf disease detection due to leaf occlusion and small disease areas.
• Existing models struggle with complex backgrounds and varying disease scales, hindering accurate identification.

Purpose of the Study:

• To propose an efficient Tomato Disease Detection Network (E-TomatoDet) for enhanced tomato leaf disease identification.
• To improve the perception of both global and local features for more robust detection.

Main Methods:

• Integrated CSWinTransformer (CSWinT) into the backbone for improved global feature capture.
• Developed a Comprehensive Multi-Kernel Module (CMKM) for multi-scale local feature learning.
• Designed a Local Feature Enhance Pyramid (LFEP) neck network to integrate multi-scale features across detection layers.

Main Results:

• E-TomatoDet achieved a mean Average Precision (mAP50) of 97.2% on a tomato leaf disease dataset, a 4.7% improvement over the baseline.
• The model surpassed the performance of the advanced YOLOv10s real-time detection network.
• Demonstrated significant improvement in detecting tomato leaf disease targets across various scales and complex backgrounds.

Conclusions:

• The proposed E-TomatoDet effectively addresses limitations in detecting tomato leaf diseases under challenging agricultural conditions.
• This research offers a viable solution for the efficient detection of vegetable pests and diseases in intelligent agriculture systems.