A lightweight rice disease identification network based on attention mechanism and dynamic convolution

Abstract

Ensuring the quality and yield of rice depends heavily on the accurate identification of early-stage rice diseases. Existing models face significant challenges in balancing lightweight requirements and precise classification of rice disease types due to the noisy background and scattered distribution of disease symptoms in real-world environments. To address the above issues, this study proposes DGLNet, a novel lightweight and highly accurate network for rice disease identification. DGLNet includes two low-complexity modules, the global attention module (GAM) and the dynamic representation module (DRM). The GAM is designed to capture key information in complex noisy scenes, thus improving the generalization ability of the model. Meanwhile, the self-developed four-dimensional flexible convolution (4D-FConv) in the DRM can dynamically generate adaptive convolutional kernel parameters from four dimensions. This allows DRM to maintain diversity among different sample inputs to enhance the model's ability to fit complex functions. As a result, DRM enhances feature representation without the need for additional network layers and channels. The proposed method achieves 99.82% and 99.71% recognition accuracy on two real plant disease datasets, outperforming current popular methods. The experimental results demonstrate that the proposed method is not only computationally lightweight, but also capable of accurately identifying rice diseases in real-life scenarios. Furthermore, this study provides robust technical support for disease identification and control and offers guidance for the implementation of agricultural intelligence and precision farming.