Recognition Of Apple Leaf Disease Based On Lightweight Neural Network

Cedar rust,black rot and scab are three common leaf diseases of apple,which seriously affect the yield and quality of apple.The identification of these diseases is one of the key technologies to ensure apple quality.However,at present,the research work of disease recognition based on machine learning is heavy,and feature extraction depends heavily on expert experience.Therefore,this dissertation proposes an apple leaf disease identification method based on deep learning.At the same time,in view of the fact that deep convolutional neural network models are generally large in scale and have many parameters,they are difficult to apply to devices with limited memory,a lightweight convolution structure based on the model is proposed in this dissertation.The main research works are as follows:(1)Based on the idea of group convolution and depth separation convolution,a lightweight model suitable for apple leaf disease recognition was designed.Taking the classical SqueezeNet model as the basic network structure,group convolution and depth separation convolution are added to its bottleneck structure(Fire module)to reduce the model parameters.In addition,the channel shuffle strategy is used to strengthen the information exchange between channels,and the fusion of deep features and shallow features is used to improve the accuracy of the model.At the same time,in order to further improve the performance of the lightweight model,this paper extends the diseased leaf data set to obtain the pre-training model based on the diseased leaf data set,and then fine tune the apple leaf disease recognition model.The experimental results show that compared with the classical SqueezeNet model,the parameters are reduced by 61%,the amount of calculation is reduced by 77%,and the accuracy is improved by 2.3%.The model recognition effect initialized by pre-training parameters is better.(2)In order to further reduce the model parameters and calculation and improve the model performance,an apple leaf disease identification model based on knowledge extraction was constructed.Firstly,in order to simplify the network structure,a more portable network structure is designed and used as a student network.Then the knowledge extraction strategy is used to improve the model,and the convolution structure with excellent performance is used as the teacher network.In order to avoid selecting teacher networks with unsatisfactory guidance effect,the idea of "further training" is proposed to improve the generalization performance of teacher network selection.The results show that the network structure parameters are only6.3% of the classical SqueezeNet model,and the accuracy is improved by 1.97%.The thought of "further training" can effectively improve the guiding effect of teachers’ network.The characteristic and innovation of the dissertation lies in: Deep learning is introduced into the apple disease identification,which solves the shortcomings of manual intervention in the early stage.In order to solve the problem that the large scale of deep learning models is difficult to apply to devices with limited memory,a lightweight apple leaf disease identification model based on deep convolutional neural networks combined with the knowledge distillation method is constructed.Which greatly reduces the parameters while improving the identification accuracy,and achieves a relative balance between the amount of parameters and the accuracy.In view of the teacher network with poor guidance,this dissertation puts forward the idea of "further training",which can effectively improve the guidance effect of the teacher model.The research in this dissertation has certain reference value and practical significance for the identification of apple leaf diseases.