DM-ConvTNet: An Effective Hybrid Model for Rice Disease Recognition in Challenging Field Backgrounds

Rice globally faces yield reduction due to various diseases and pests, posing a significant threat to food security. Accurate identification of rice diseases and pests in real field conditions is challenging owing to the chaotic background and varying sizes of affected areas in the captured images. In this study, we propose a lightweight hybrid model named DM-ConvTNet, which integrates convolution and transformer structures to address this issue. Our model incorporates a dual-branch joint attention mechanism, aiding the convolutional layers in mitigating interference from complex backgrounds during feature extraction. Then, the linear vision transformer module focuses on the global disease information in the images. Finally, a lightweight version of the Inception module is employed for multi-scale feature extraction of high-level semantic information. Experimental results demonstrate that DM-ConvTNet excels in balancing model computational requirements and recognition performance. On two real-field rice disease datasets, DM-ConvTNet achieves impressive recognition accuracy, outperforming other state-of-the-art lightweight models. These results validate the effectiveness of our approach, highlighting its practical utility in addressing the challenges of rice disease and pest identification in complex field environments.