Application Research On Plant Disease Identification Based On Depthwise Separable Convolution

The traditional methods of identifying plant diseases mostly rely on experience and manually determining the disease category,which has poor identification efficiency and effect.As more and more scholars apply machine learning methods to agriculture,the efficiency of identifying plant diseases has been greatly improved.Still,the traditional machine learning methods need to select and extract features artificially according to experience,and the recognition efficiency still needs to be improved.Although the deep learning recognition method can automatically extract the feature information,reduce the labor cost,and improve the accuracy,at the same time,the existing plant disease recognition model based on deep learning also has some problems,such as a large number of parameters,bloated model,slow recognition speed and recognition accuracy to be improved.Given this,this study takes the open data set Plant Village as the research object,combined with machine learning,deep learning,transfer learning,and other related knowledge,uses and improves the convolution neural network structure,and applies it to the research of plant disease identification through iterative training.The specific work of this paper is as follows:（1）Aiming at the problems of cumbersome work,heavy workload,and long time-consuming in identifying plant diseases by traditional digital image processing methods,this paper adopts three convolution neural networks: ShuffleNet,MobileNetV3,and Efficient Net,through a large number of training and identification verification of plant disease data sets,and through a large amount of preprocessing work on the data sets in advance,The super parameter settings of the experimental network model are compared to improve the recognition effect.（2）Aiming at the problems of excessive network parameters,a large amount of calculation,long training time and convergence time,and an unbalanced distribution of categories in disease data set,through a series of strategies such as replacement optimizer,loss function,network structure optimization,the introduction of attention mechanism and transfer learning,the overall performance of the network model is improved,and an improved model based on ShuffleNet: FL-ShuffleNet is proposed.（3）Integrating the above research contents,taking FL-ShuffleNet as the basic network of plant disease identification,a real-time plant disease detection system based on Android is developed.Users can upload disease images without geographical restrictions and obtain disease diagnosis information in time by installing an APP on mobile devices,to improve the effect of plant disease identification,diagnosis,and prevention and promote professional agricultural knowledge.