A multi-scale parallel weighted fusion dynamic attention method for citrus leaf disease recognitions.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Citrus leaf disease detection faces challenges like low accuracy due to leaf occlusion, target loss, and complex backgrounds.
• Existing methods struggle to effectively extract and fuse disease features in diverse environmental conditions.

Purpose of the Study:

• To develop an advanced real-time detection transformer for improved citrus leaf disease identification.
• To enhance accuracy and efficiency in disease detection within complex agricultural settings.

Main Methods:

• Proposed DBG-DETR method incorporating DMGF-ResNet18 for feature extraction and GSDT for focusing on deep features.
• Utilized multiscale parallel depthwise separable convolutions and dynamic gating with Top-K sparse attention.
• Implemented a bi-directional dense feature fusion module (BDFF) for effective shallow and deep feature interaction.

Main Results:

• DBG-DETR demonstrated significant improvements over the baseline model, with increases in P (3.31%), mAP (3.40%), mmAP (4.11%), R (3.89%), and F1 (3.59%).
• The model achieved a reduction of 3.78 MB in parameters, indicating improved efficiency.
• Experimental results confirmed enhanced disease detection performance in complex environments.

Conclusions:

• The proposed DBG-DETR method effectively addresses challenges in citrus leaf disease detection.
• The study provides a reliable technical foundation for intelligent citrus orchard management systems.
• DBG-DETR offers a promising solution for accurate and efficient plant disease identification.