Research On Crop Disease Image Recognition Based On EfficientNetV2

The occurrence of crop diseases has seriously affected the yield and quality of crops,resulting in huge economic losses.In order to prevent large-scale crop yield loss,it is crucial to accurately identify the type of crop disease and the severity of a certain disease.With the rapid development of information technology,computer technology is widely used in crop disease identification research,which also promotes the advancement of agricultural modernization.Traditional crop disease image recognition methods have problems such as low accuracy and cumbersome steps.With the development of convolutional neural networks in the field of image recognition,it also provides a new idea for crop disease image recognition.In order to accurately and timely identify the types of crop diseases,this paper combines convolutional neural networks with crop disease image recognition tasks,and uses 15 types of crop disease images disclosed in Plant Village to train five types of convolutions: Alex Net,VGG16,Res Net50,Google Net and EfficientNetV2 Neural Networks.Because most convolutional neural networks have the disadvantage of large number of parameters,the number of parameters will directly lead to the increase of model calculation volume（FLOPs）,and at the same time,it will increase the demand for hardware equipment,which makes the model difficult to deploy in practical applications.Therefore,on the basis of ensuring the accuracy of model recognition,this paper further considers the amount of parameters and calculation of the model,and selects the EfficientNetV2 lightweight network as the baseline model of this paper to deal with the two tasks of this paper: recognition of multiple diseases of crops and disease severity recognition.In the first task,the EfficientNetV2 network has the advantage of automatically scaling the depth,width and image resolution of the network.In order to further improve the accuracy of the EfficientNetV2 model for identifying crop diseases,this paper uses an online label smoothing algorithm（Online Label Smoothing,OLS）combined with the loss function to optimize the EfficientNetV2 model.The OLS algorithm can use the prediction statistics of the target category to dynamically adjust the generation of category soft labels,so that the optimized EfficientNetV2 model can improve the accuracy of identifying crop diseases.Experiments show that the optimized EfficientNetV2 achieves the performance of other models with a large number of parameters,and the accuracy rate of the test set reaches 98.8%,which is 2.5% higher than that before optimization.In the second task,this paper further studies the severity identification of single crop disease based on EfficientNetV2 network.In this paper,potato early blight images are divided into four severities: healthy,early stage,middle stage,and late stage,and the divided dataset is augmented with data augmentation.Due to the characteristics of large intra-class feature similarity and small inter-class feature variance in fine-grained crop disease images,it is difficult to classify categories with fuzzy boundaries.To solve this problem,this paper starts from the intra-class relationship and the inter-class relationship,and uses the cross-entropy loss function and the center loss function in combination to compact the intra-class sample features while ensuring the separability of the inter-class features.Experiments show that the use of weighted loss function can make it easier for the optimized model to distinguish categories with fuzzy boundaries in potato early blight,and the recognition accuracy on the test set reaches 97.78%,which is 2.18% higher than that of the basic model.