Study on lightweight rice blast detection method based on improved YOLOv8.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Timely detection of rice diseases is crucial to prevent significant yield reduction and economic losses for farmers.
• Existing methods often struggle with accuracy and model efficiency for real-time applications.

Purpose of the Study:

• To develop a lightweight and accurate rice disease detection model.
• To improve feature extraction, model robustness, and bounding box regression for enhanced performance.

Main Methods:

• An improved YOLOv8 architecture incorporating a full-dimensional dynamic convolution (ODConv) module.
• Integration of the WIoU (weighted interpolation of sequential evidence for intersection over union) mechanism for bounding box loss optimization.
• Utilizing a high-resolution detector head to improve small object detection and reduce network parameters.

Main Results:

• Achieved a 66.6% reduction in parameters and a 61.9% reduction in model size compared to YOLOv8n.
• Demonstrated superior mean average precision (mAP) over established models like Faster R-CNN, YOLOv5s, YOLOv6n, YOLOv7-tiny, and YOLOv8n.
• Significant improvements in detection performance, indicating enhanced accuracy and efficiency.

Conclusions:

• The YOLOv8-OW model offers an effective, lightweight solution for real-time rice disease detection.
• Suitable for deployment on resource-limited mobile devices, supporting precision agriculture.
• Contributes to advancing accurate and timely disease management for farmers.