Wheat disease detection is crucial for disease diagnosis, pesticide application optimization, disease control, and wheat yield and quality improvement. However, the detection of wheat diseases is difficult due to their various types. Detecting wheat diseases in complex fields is also challenging. Traditional models are difficult to apply to mobile devices because they have large parameters, and high computation and resource requirements. To address these issues, this paper combines the residual module and the inception module to construct a lightweight multiscale CNN model, which introduces the CBAM and ECA modules into the residual block, enhances the model’s attention to diseases, and reduces the influence of complex backgrounds on disease recognition. The proposed method has an accuracy rate of 98.7% on the test dataset, which is higher than classic convolutional neural networks such as AlexNet, VGG16, and InceptionresnetV2 and lightweight models such as MobileNetV3 and EfficientNetb0. The proposed model has superior performance and can be applied to mobile terminals to quickly identify wheat diseases.

Wheat production safety is facing serious challenges because wheat yellow rust is a worldwide disease. Wheat yellow rust may have no obvious external manifestations in the early stage, and it is difficult to detect whether it is infected, but in the middle and late stages of onset, the symptoms of the disease are obvious, though the severity is difficult to distinguish. A traditional deep learning network model has a large number of parameters, a large amount of calculation, a long time for model training, and high resource consumption, making it difficult to transplant to mobile and edge terminals. To address the above issues, this study proposes an optimized GhostNetV2 approach. First, to increase communication between groups, a channel rearrangement operation is performed on the output of the Ghost module. Then, the first five G-bneck layers of the source model GhostNetV2 are replaced with Fused-MBConv to accelerate model training. Finally, to further improve the model’s identification of diseases, the source attention mechanism SE is replaced by ECA. After experimental comparison, the improved algorithm shortens the training time by 37.49%, and the accuracy rate reaches 95.44%, which is 2.24% higher than the GhostNetV2 algorithm. The detection accuracy and speed have major improvements compared with other lightweight model algorithms.