Research On Recognition Method Of Crop Disease Based On Lightweight Neural Network

The problem of crop diseases is closely related to people’s production and is one of the main reasons for the economic losses of growers.The focus of disease prevention and control is to determine the type of disease in time and control the condition before causing serious losses.The current disease recognition method based on the convolutional neural network has the advantage of low recognition error rate,but it also has the problems of numerous parameters,a large amount of calculation,and complexity,which affect the practicability of the method.However,most of the existing lightweight disease recognition networks fail to integrate disease feature analysis,and their performance needs to be improved,and they are not suitable for disease images taken in natural scenes,and have certain application limitations.In response to these problems,this research constructs a lightweight crop disease recognition network,which is suitable for disease images taken in natural scenes and disease images with a simple background,and designed and implemented a crop disease recognition system.This research has strong practicability and is of great significance to the prevention and control of crop diseases and the improvement of crop quality and yield.The specific research content is as follows:(1)Crop disease recognition based on simple background images.Aiming at the problems of crop disease recognition networks based on simple background images with numerous parameters,a large amount of calculation and complexity,a lightweight crop disease recognition network — split-fuse net is proposed.The network reduces the recognition error rate by fusing the disease characteristics of different scales.Reduce parameters and calculations through a depthwise separable convolution and optimized network structure.Experiments on the Plant Village dataset show that,compared to Mobile Net V3 and Shuffle Net V2,the split-fuse net has a smaller amount of calculation,fewer parameters,and a lower recognition error rate.(2)Crop disease recognition based on natural scene images.The crop disease images taken in natural scenes often contain complex backgrounds and disease features are more likely to appear anywhere in the image,which affects the recognition accuracy of the convolutional neural network.To solve this problem,the split-fuse net for nature suitable for crop disease images taken in natural scenes is proposed.The max-pooling and efficient channel attention(ECA)module are integrated into the original network to reduce the interference of complex background and other information and reduce the excessive dependence of the network on the feature location.Through data set expansion,avoid model overfitting.Experiments on the natural scene crop disease data set(including 6 crops and 14diseases)show that the split-fuse net for nature has a small amount of calculation,fewer parameters,and a low recognition error rate,which is suitable for crop disease images taken in natural scenes.(3)Implementation of crop disease recognition system.In order to promote the application of disease recognition methods,this research designed and developed a farming disease recognition system.The system is based on the We Chat applet and contains three functional modules: disease knowledge,disease recognition,and disease addition.The disease knowledge module mainly pushes news,prevention methods,and other articles related to crop diseases.The disease recognition module is based on split-fuse net / for nature,supports a variety of crop diseases,and has a low recognition error rate.At the same time,combined with tensorflow.js,the model is deployed on the mobile terminal.The disease-adding module can upload other disease images to promote system expansion.