A Multi-Scale Perception-Enhanced Lightweight Network with Knowledge Distillation for Rice Leaf Disease Detection.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Rice diseases like blast and bacterial leaf blight threaten global food security.
• Current disease detection methods lack efficiency and generalization, failing to capture variations across growth stages.

Purpose of the Study:

• To develop a lightweight and efficient rice leaf disease detection model.
• To improve detection accuracy and generalization capability for complex environmental conditions.

Main Methods:

• Proposed a lightweight knowledge distillation YOLO (LWKD-YOLO) model, replacing YOLOv8 convolutional layers with ADown modules.
• Introduced a lightweight shared re-parameterizable convolutional detection head (LSRP-Head) with group normalization RepConv.
• Utilized YOLOv8x as a teacher model for inter-channel correlation knowledge distillation.

Main Results:

• LWKD-YOLO increased mAP@50 by 1.4% compared to the baseline model.
• The model achieved a reduction of 1.3M parameters and 3.1G FLOPs.
• Demonstrated enhanced ability to learn complex rice leaf disease features.

Conclusions:

• The proposed LWKD-YOLO model enables efficient and accurate rice leaf disease detection in complex environments.
• The model holds significant economic and practical value for sustainable agriculture and food security.