An Improved Multi-Scale Feature Extraction Network for Rice Disease and Pest Recognition

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Rice production faces significant yield losses due to pests and diseases.
• Accurate identification of rice pests and diseases is essential for effective management strategies.
• Current image recognition technologies for rice pest and disease identification are limited by data scarcity and low accuracy.

Purpose of the Study:

• To address the limitations in current rice pest and disease identification methods.
• To develop a more accurate and efficient image recognition model for rice pests and diseases.
• To construct and expand a comprehensive rice pest and disease dataset (RPDD).

Main Methods:

• Construction of the Rice Pest and Disease Dataset (RPDD) with data augmentation.
• Proposal of a Lightweight Multi-scale Feature Extraction Network (LMN) based on ResNet and attention mechanisms.
• Extraction of multi-scale features at a finer granularity for improved classification.

Main Results:

• The proposed LMN model achieved an average classification accuracy of 95.38% on the RPDD.
• The LMN model obtained an F1-Score of 94.5%, demonstrating high classification performance.
• The LMN model has a small parameter size (1.4 M) and low computational cost (1.65 G FLOPs).

Conclusions:

• The LMN model significantly outperforms the baseline ResNet model in rice pest and disease classification.
• The LMN model offers a lightweight and accurate solution for agricultural image recognition tasks.
• This research contributes to advancing automated pest and disease identification in rice production.