GCPDFFNet: Small Object Detection for Rice Blast Recognition.

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision
• Machine Learning

Background:

• Rice blast disease poses a significant threat to global rice production and food security.
• Accurate and timely detection of rice blast is essential for effective disease management and yield preservation.
• Existing detection methods may struggle with the small size and varied appearance of rice blast lesions.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for the early and accurate detection of rice blast disease in situ.
• To address the challenge of recognizing small, irregularly shaped rice blast lesions in field images.
• To improve the efficiency and accuracy of automated rice blast detection systems.

Main Methods:

• Construction of a specialized rice blast dataset featuring diverse lesion characteristics (shape, size, color).
• Proposal of the Global Context-based Parallel Differentiation Feature Fusion Network (GCPDFFNet), incorporating global context and parallel differentiation feature fusion modules.
• Introduction of the SCYLLA normalized Wasserstein distance loss function to enhance model convergence and detection accuracy.

Main Results:

• The GCPDFFNet model achieved a significant improvement in mean average precision (mAP), increasing from 83.6% to 95.4% compared to the baseline CenterNet.
• The model maintained a high inference speed of 122.1 frames per second, meeting real-time detection requirements.
• Experimental validation confirmed the superior performance of GCPDFFNet in accurately identifying rice blast lesions.

Conclusions:

• The proposed GCPDFFNet model demonstrates high accuracy and efficiency for in situ rice blast disease detection.
• The novel network architecture and loss function effectively address the challenges of small lesion recognition.
• This advancement holds potential for practical application in precision agriculture and crop disease management systems.