Research And Implementation Of Wheat Disease Detection Method Based On Edge Computing

Wheat is one of the most important food crops in our country.It is very important to increase and stabilize the yield of wheat.However,global climate change leads to frequent wheat diseases,which seriously affects wheat yield.The rapid development of artificial intelligence technologies such as machine learning has injected new impetus for the intelligent development of agriculture,and also provided technical guarantee for better solving the problem of wheat diseases and ensuring national food security.In this era of the Internet of Things,the number of devices grows at an explosive speed,and a large amount of data is generated every moment,which makes the existing cloud computing model face a severe test and can no longer adapt to the current computing needs.The main idea of edge computing is to make data be analyzed and processed near the data source,so it is widely used.In addition,with the rapid development of deep learning and the combination of artificial intelligence and edge computing,the requirement for the intelligent development of edge devices is getting higher and higher.However,edge devices often cannot meet the requirements of high storage and high energy consumption of deep learning models,so compressing convolutional neural network model becomes a major challenge.This paper takes the research of wheat diseases detection model as the breakthrough point and takes image processing and model compression to establish a lightweight wheat disease detection model that can be deployed on edge device,which effectively reduces the false detection rate and missed detection rate,and greatly improves the early detection efficiency of wheat diseases.It is convenient for farmers to take timely prevention and control measures and effectively reduce the loss of wheat yield.The main work and research contents of this paper are as follows:(1)A wheat disease detection model combined with Coordinate Attention(CA)is proposed,which effectively improves the problem of missed and false detection in the wheat diseases detection model.Due to the fact that Yangzhou is located in Lixiahe River region of Jiangsu Province,the main wheat disease is Fusarium Head Blight.Therefore,the research object of this paper is Fusarium Head Blight,and the established wheat disease detection model is also applicable to other diseases.Firstly,the dataset of Fusarium Head Blight was collected,and the wheat images were enhanced by data augmentation operations such as random rotation.Then,LabelImg software was used to label the disease in the wheat image.In this paper,YOLOv4 was selected as the basic model,and the dilated convolution was fused in the feature extraction network to obtain more semantic information and location information.and the Coordinate Attention module was added in the enhanced feature extraction network.so that shallow feature information could be fully extracted.In order to solve the problem of unbalanced positive and negative samples.Focal loss classification loss function was introduced.In addition,Class Activation Mapping(CAM)was used to visually observe whether the model paid more attention to the important information after the addition of the attention module.The experimental results showed that the detection accuracy of the improved model CoordAtt-YOLOv4 for wheat diseases reached 91.30%after fusion of the dilated convolution and the addition of attention mechanism,which was about 3.4%higher than the original model.(2)A lightweight wheat disease detection model based on deep separable convolution is proposed,which effectively improves the problems of large memory and high energy consumption of deep learning model and difficult deployment to edge device.Since the memory space of YOLOv4 model is large and the storage space of edge device is limited,it is difficult to deploy it directly to edge device.Therefore,this paper carried out lightweight processing on the basis of the previous wheat disease detection model combined with attention mechanism.Firstly,Mosaic data enhancement operation was used to replace conventional data enhancement operations in wheat disease images to increase the amount of training data.Then.based on the improved model CoordAtt-YOLOv4,MobileNet series networks were used to replace the backbone feature extraction network.In order to further reduce the number of parameters and computation,the common convolution of enhanced feature extraction network was replaced by deep separable convolution.The experimental results showed that compared with other advanced models,the memory size of the improved model MobileNetv3-YOLOv4CA was reduced to 20%of the original model,the accuracy of model detection reached 94.69%,and the average detection speed was 0.111 seconds,which was more suitable for deployment on edge device.(3)A wheat disease detection system based on Jetson Xavier NX is constructed,which can achieve real-time monitoring of wheat diseases.After placing the edge device in the wheat field,and receiving the detection instruction sent by the web client user,Jetson Xavier NX first collected and stored the image,performed data processing near the data end,and then transmitted the processing results to the server database in the corresponding format,and fed back the results to the client user for visualization in the form of charts.In addition,the client can obtain an overview of common wheat diseases and prevention methods,and can also detect diseases that exist locally.