Research On Recognition Methods Of Corn Leaf Diseases Based On Edge Intelligence

Corn is the main food crop and livestock feed,the demand for corn in various countries has been large.This means our country also need urgently solve to increase the output of corn and drive the economic development of farmers.The increasing variety of corn diseases is the main reason restricting the increase in yield.Therefore,using computer technology to help farmers identify corn diseases is one of the key researches of scholars.As an emerging technology,Neural Networks image recognition technology has gradually become the mainstream of image recognition.However,the training and inference of the Neural Networks model requires high performance of the computer,it is generally completed in the cloud server.With the development of Cloud Computing,the problem of centralized storage and processing of data has gradually surfaced.On the one hand,the rapid growth of data has increased the burden of cloud servers;On the other hand,the cloud server is far from the user and data source,it cannot guarantee real-time requirements of users.The emergence of Edge Computing solves these problems.The essence of Edge Computing is that pushes computing tasks and data storage near users,and provides intelligent services nearby.The rapid identification of corn diseases can not only reduce the use of pesticides and costs,but also increase the yield of corn.However,the corn leaf diseases identification technology based on Neural Networks technology is computationally intensive and time-consuming.So its real-time performance needs to be improved.The paper studies the methods of using Neural Network technology to identify corn leaf diseases under the framework of edge computing,and mainly completes the following four tasks:(1)The strategies based on edge intelligent image recognition is proposed.The Terminal-Edge collaborative strategy divides image reasoning into two parts to execute.Part of the program is executed on the terminal device,and then the inference result is executed at the edge node;The Terminal-Cloud collaboration strategy is similar to the Terminal-Edge collaboration strategy.It also divides the calculation into two parts,but the latter part of the reasoning needs to be carried out in the cloud;The Terminal-Edge-Cloud collaborative strategy uses cloud computing and edge computing.Data preprocessing is done on mobile devices,and then inferences are completed on edge node and cloud in sequence.The collaboration strategies take advantage of that edge devices being closer to the data source and stronger computing processing capabilities of the cloud server,thereby improving the performance of the image identification system.(2)The method to improve the accuracy of identifying corn leaf diseases was studied.Because the Alex Net model has a dropout layer,which reduces the overfitting of the model.We chose the Alex Net model.The hyperparameters are adjusted and the accuracy is increased to91.83%.At the same time,it is found that overlapping pooling is not suitable for the data set.The pooling layer is changed to Non-overlapping pooling,and the pooling combination is Max-Ave-Max.Finally,the average accuracy is increased to 96.75%,which also reduces the loss rate.(3)The calculation partition of the Alex Net model is studied.The paper sets the exit point according to the amount of data generated and the delay time by each layer calculation.The exit point is set as between Pooling5 and FC6.The data that needs to be transmitted after the exit point is small,and the amount of calculation required behind is large,which is suitable for execution at the edge or cloud center.(4)The performance of collaborative strategies based on edge intelligence is analyzed.The accuracy of inferring images under these three strategies is not different.But the delay contrast is very big.The total delay of recognizing images in the cloud center is 606 ms,and the delay of recognizing images under the Terminal-Edge collaborative strategy is less than half of the inference of the cloud center.The reduced delay time under the Terminal-Cloud collaboration strategy is much less.The Terminal-Edge-Cloud strategy is not optimal for the results of this experiment,but it is a better choice for Neural Networks with more complex calculations.For energy consumption,it is found that the execution on the GPU is much smaller than that on the CPU.According to the analysis,it is found that the collaborative strategies based on edge intelligence improves the performance of image recognition.