A Copy Paste and Semantic Segmentation-Based Approach for the Classification and Assessment of Significant Rice...

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate rice disease segmentation and severity assessment are crucial for early diagnosis, intelligent monitoring, and effective pest control.
• Deep learning models offer high accuracy in disease detection but require extensive training data.
• Existing methods often struggle with data limitations and model efficiency.

Purpose of the Study:

• To propose a lightweight network for accurate rice disease region segmentation and severity assessment.
• To introduce a novel data augmentation method to enhance training dataset diversity and model generalization.
• To develop an efficient deep learning model for real-time rice disease identification and management.

Main Methods:

• A dataset of 450 rice disease images (bacterial blight, blast, brown spot) was curated from open-source datasets.
• A data augmentation technique, Rice Leaf Disease Copy Paste (RLDCP), was developed to increase sample diversity.
• A lightweight semantic segmentation model, RSegformer, was proposed, incorporating a lightweight backbone, attention mechanism, and modified upsampling.

Main Results:

• The RLDCP data augmentation method significantly improved the accuracy and generalization performance of semantic segmentation models.
• RLDCP improved Mean Intersection over Union (MIoU) by approximately 5% with a dataset twice the original size compared to traditional augmentation.
• The RSegformer model achieved an 85.38% MIoU with a compact model size of 14.36 M.

Conclusions:

• The proposed RLDCP method effectively enhances semantic segmentation model performance with limited data.
• RSegformer provides an efficient and accurate solution for rice disease segmentation and severity assessment.
• This approach offers a valuable reference for timely and effective rice disease control strategies.