Research On Crop Disease Identification Via Lightweight CNN

The problem of crop diseases is one of the main problems faced in agricultural development and one of the main factors for the loss of growers.The main point of disease control is to determine the type of disease in the early stage and control it to reduce agricultural economic losses.However,there are many types of crop diseases,and there are similarities and differences between species,which are easy to be confused in the identification process,and the corresponding control measures for different types of diseases are also different.Therefore,real-time and accurate detection of crop disease types is effective.important prerequisite for disease control.However,most convolutional neural network(CNN)models have the characteristics of many parameters and large volumes.Such network models have high requirements on memory and computing resources,and the inference calculation takes too long,which is difficult to apply to small mobile devices.The recognition accuracy of the lightweight convolutional neural network is not high enough.This paper uses its strengths to make up for its shortcomings to improve the recognition efficiency of the lightweight model on crop datasets.The main research contents and results are as follows:(1)In order to improve the convergence rate of the model,a transfer learning method is introduced,and the parameters of the model that have been pre-trained on the Image Net dataset are transferred to the corresponding network and its fully connected layer is modified.To a certain extent,it can improve the convergence speed of the model and further improve the generalization ability of the model.(2)In view of the characteristics of large volume,many parameters,large amount of calculation and lightweight model of classical complex model,small volume of parameters and small amount of calculation,but limited identification efficiency,a lightweight model for detecting crop leaf diseases is proposed.The model combines lightweight CNN,transfer learning and knowledge distillation methods,and selects the res Net101 model with higher accuracy as the teacher model,and the smaller and more stable Mobile Net V3 model as the student model for training.The knowledge distillation method is used to enable the student model to learn the knowledge of the teacher model,which improves the overall efficiency of the student model to a certain extent.At the same time,in order to make the features trained by the model more discriminative,the loss function used during training has been improved,and the Center loss function and the Softmax loss function are used in combination.The Center loss function is assisted by the Softmax loss function for training to ensure that the leaf image features of the same disease species are as compact as possible,while the leaf image features of different species remain relatively discrete,making the identification method and results of crop diseases more robust.Robustness,to a certain extent,it solves the problem of insufficient accuracy of lightweight models.Then we discuss the effect of different hyperparameters on model accuracy.Finally,the features are visualized,so that the extracted features can be displayed more intuitively,and it is determined that the classification of disease images is based on the disease area.The results show that for the Plant Village dataset,the accuracy of the improved model on the test set is increased by 1.2%,and the recognition time on the GPU is reduced by 0.032 s.This approach preserves the lightness of the student model and improves its performance,providing a theoretical basis for its subsequent deployment to mobile devices.(3)The disease identification system is built in the Android Studio environment to test the performance of the model on Android phones.The system can realize the functions of reading pictures in the album for identification and photo identification.The inference time is about 62 ms,and the memory usage of the system is only 38 M,indicating that the recognition efficiency of the model is high and the storage space occupied by the system is less.In summary,this study has strong practicability,can improve the recognition efficiency of the model to a certain extent,and can be applied to mobile phones,indicating that the method used in this paper has certain significance for the identification and prevention of crop diseases.