ConTransformer: Integrating Local and Global Features for Accurate Identification of Rice Leaf Diseases

Blast, brown spot, leaf scald, bacterial leaf blight, and leaf smut severely threaten rice yield and quality. To accurately identify these diseases, this paper proposes a dual-channel fusion model named ConTransformer, which integrates ConvNeXt for fine-grained lesion feature extraction and Vision Transformer for global pattern modeling. In the ConvNeXt branch, lightweight convolutions and Global Response Normalization are introduced to capture detailed lesion textures and alleviate the limitations of conventional CNNs in long-range dependency modeling. In the Vision Transformer branch, a Swish-gated MLP is incorporated to strengthen global feature representation and suppress irrelevant noise, addressing the insufficient local lesion sensitivity of standard ViT-based methods. Finally, a feature fusion module combines multi-scale representations through concatenation and attention mechanisms. Experiments on a self-built seven-class rice disease dataset containing 14,280 images demonstrate that ConTransformer achieves 99.07% accuracy, with precision, recall, and F1-score reaching 99.01%, 98.87%, and 98.93%, respectively. Compared with MobileNetV3, Swin Transformer, LGNet, DB-CGANNet, Efficient-ViT B0Net, LocalViT, and TinyViT, ConTransformer achieves the highest accuracy, exceeding them by 1.75%, 0.72%, 10.33%, 0.65%, 0.94%, 0.87%, and 1.44%, respectively. Moreover, cross-crop transfer learning on a soybean disease dataset obtains 92.78% accuracy, further verifying the model’s generalization ability. These results indicate that ConTransformer provides an effective and robust solution for rice disease identification.