Corn leaf disease recognition based on improved EfficientNet

Abstract

Abstract Maize leaf disease seriously affects maize yield, a maize leaf disease identification model with an improved lightweight network EfficientNet was proposed in this study. First, the model replaces the SENet module in the MBConv module with the CBAM module, so that the model not only focuses on the correlation between the channels but also adaptively learns the attentional weight of each spatial location. Furthermore, a multi‐scale feature fusion layer based on residual connection is introduced to extract more comprehensive and richer disease features at different scales. Finally, by introducing the double pooling method, the overall feature distribution is smoothed while highlighting the important disease features. After the three improvements, the model's recognition accuracy on the test set increased by 2.34%, 2.16%, and 0.97%, respectively, and the improved model achieved an average recognition accuracy of 98.32%, an average precision of 98.29%, and an average recall of 98.25%. The experimental results compared with other models show that the average recognition accuracy of the proposed model is 5.23%, 3.68%, 1.99%, 1.79%, and 3.2% higher than ResNet34, DenseNet121, MobileNet V2, SqueezeNet, and EfficientNet B0, respectively. Activation heat maps show that the improved model can effectively suppress background interference.

Topics & Concepts