AF-RT-DETR: Adaptive cross-scale feature interaction for real-time plant disease detection in complex field...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate plant disease identification is crucial for food security.
• Existing methods struggle with illumination variations, occlusion, and scale diversity, limiting real-time detection.
• Challenges in real-time plant disease detection impact agricultural productivity.

Purpose of the Study:

• To develop an improved real-time plant disease detection model.
• To enhance accuracy and robustness in complex agricultural environments.
• To address limitations of current plant disease identification approaches.

Main Methods:

• Proposed an improved RT-DETRv2-based model named AF-RT-DETR.
• Introduced a Bidirectional Cross Gate (BCG) module for enhanced feature representation.
• Integrated a Dynamic Channel Shift (DCS) module and an improved Scale-aware Multi-level Loss (SML) for better performance.

Main Results:

• Achieved mAP50 of 93.6% and mAP50:95 of 67.2% on the Plant-Disease dataset.
• Outperformed the baseline model by 5.1% and 4.5% in key metrics.
• Demonstrated robust performance and adaptability across multiple crops and field conditions.

Conclusions:

• AF-RT-DETR effectively enables real-time plant disease detection.
• The model shows significant improvements in accuracy and robustness.
• The proposed methods are suitable for complex, real-world agricultural settings.