DSCRANet : A Novel Potato Leaf Disease Recognition Model Based on Multi‐Scale Residual Attention Blocks and Depthwise Separable Convolution in Complex Backgrounds

Crop diseases seriously affect the stable development of the agricultural economy. Accurately identifying common crop diseases is crucial for improving the production quality of crops. However, due to the limitations of model performance, traditional deep learning algorithms cannot effectively diagnose potato leaf diseases in complex scenarios. Therefore, this study has constructed a novel residual model called “Depthwise Separable Convolution Residual Attention Network” ( DSCRANet ) to improve the classification accuracy of common potato diseases. Firstly, to better capture global disease features, a new multi‐scale residual block was proposed. This module uses multiple parallel convolutional paths to efficiently extract features from the input samples, and combines depthwise separable convolution operations to reduce the computational complexity of the model. Secondly, a dual‐path enhanced squeeze and excitation ( DESE ) attention mechanism was proposed and embedded into the multi‐scale residual block to further improve the model's performance. This attention module performs a unique dot‐product operation on the input features to increase the difference between disease information and other irrelevant information. To achieve better classification performance, the RFB module and ELU activation function were also introduced into the model. The classification results of the potato disease dataset show that DSCRANet achieved an optimal classification accuracy of 97.07%, while significantly reducing the training time. Moreover, its precision, recall, and F1 ‐score reached 96.59%, 97.13%, and 96.86% respectively, outperforming other advanced recognition networks in terms of performance. The Grad‐ CAM visualisation results indicate that the DSCRANet algorithm not only accurately locates the disease information but also has a larger receptive field. These findings demonstrate the superiority of the DSCRANet algorithm in terms of classification performance and feature extraction effectiveness. Therefore, the DSCRANet algorithm can be regarded as a better choice for diagnosing potato diseases in complex environments.