Leaf-DETR: Progressive adaptive network with lower matching cost for dense leaves detection

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Precise leaf monitoring is vital for smart agriculture, enabling accurate assessment of photosynthesis and plant growth.
• Existing dense leaf detection methods struggle with occlusion, overlap, and high density in real-world agricultural settings.
• Challenges include incomplete feature extraction and difficult network convergence due to dense leaf scenarios.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate and efficient dense leaf detection.
• To address limitations of current methods in complex agricultural environments.
• To improve downstream applications like phenotypic trait extraction and yield estimation.

Main Methods:

• Proposed Leaf-DETR framework utilizing Progressive Feature Fusion Pyramid Network (P-FPN) and Crowded Query Refinement Strategy (CQR).
• Constructed the largest dense leaf detection dataset with 1696 images and 85,375 annotations.
• P-FPN enhances feature interaction via multi-stage fusion and Adaptive Feature Aggregation (AFA).
• CQR strategy improves efficiency through query culling and one-to-many matching.

Main Results:

• Leaf-DETR achieved a 1% improvement in mAP@50 and a 1.4% improvement in AR@300 over baseline models on the custom dataset.
• Demonstrated superior performance compared to existing detection methods.
• Exhibited fast training convergence and strong generalization across different crop types and field conditions.

Conclusions:

• Leaf-DETR effectively addresses challenges in dense leaf detection, offering practical value for complex agricultural scenarios.
• The framework shows significant potential for enhancing crop surveillance and management in smart agriculture.
• The developed dataset and model contribute to advancing automated plant monitoring technologies.