Research And Implementation Of Rice Disease Edge Intelligent Early Warning Based On Deep Learning

As an important crop in China,rice is widely planted in our country and is an important guarantee for people’s food and clothing.With the occurrence of extreme climates,the problem of rice diseases has become more and more prominent,seriously threatening our country’s food security,and becoming an obstacle to increasing food production.The rapid development of artificial intelligence technologies such as machine learning and deep learning has injected new vitality into the agricultural modernization system,and they provided a technical basis for better solving the problem of rice diseases and reducing food losses.With the proliferation of IoT(Internet of Things)devices and the geometric growth of data,real-time data transmission has become a new challenge today.Although traditional cloud computing has abundant computing resources,a large number of computing tasks easily lead to increased delay,and the real-time requirements of tasks cannot be met.The emergence of edge computing is expected to alleviate the problem of data explosion,and the edge and distribution of IoT devices has become a new development direction in smart agriculture.Based on the above-mentioned status and needs of our country’s modern agricultural development,how to use advanced science and technology such as deep learning and edge computing to solve the problem of yield loss caused by rice diseases has become a proposition worthy of in-depth discussion,this paper mainly studies the following contents:(1)An improved CNN(Convolutional Neural Network)model combining batch normalization layer and regularization norm penalty is proposed to improve the problem of slow convergence of the CNN model effectively in the rice disease identification task.Firstly,sheath blight,rice blast,rice smut,bacterial blight,brown spot and healthy rice are selected as the target categories to create the corresponding rice image dataset.Annotating images provides the basis for training neural network models.After preprocessing the image,such as data enhancement,select classic convolutional neural networks such as ResNet101,DenseNet121,MobileNetV2 to train the corresponding models.After adding the batch normalization layer and the regularization norm penalty,the training is performed again.The experimental results show that the improved model has advantages in terms of faster convergence.For the consideration of model edge deployment,the lightweight and improved MobileNetV2 model is selected as the initial model,and the model test accuracy rate reaches 89.17%.(2)A lightweight MobileNetV2 model combining parameter pruning and parameter quantization is proposed to solve the problem of limited storage and computing power in the edge deployment of the model.This paper firstly performs channel pruning on the MobileNetV2 model,and it is easy to implement and flexible.Then perform QAT(Quantize Aware Training)quantization to quantize the model data type from FP32 asymmetrically to INT8.The model size is further reduced to accommodate the limited computing power and storage environment of edge devices.The experimental results show that the model lightweight method combining parameter pruning and parameter quantization reduces the memory size of the model to 1/12 of the original,and shortens the inference time by 77.7%.The model accuracy rate reaches 87.36%.The lightweight model sacrifices some accuracy,but is more suitable for deployment in edge environments.(3)An intelligent early warning system for rice diseases at the edge based on Jetson Xavier NX was built to realize real-time monitoring of rice diseases.The edge device Jetson Xavier NX is placed in the rice field as the edge side,and waits for the detection command from the web client.Jetson Xavier NX first collects and stores images,performs data processing locally,and then wirelessly transmits the processing results to the cloud database in the corresponding format.The web client displays the result feedback in real time,and continuously accesses the database to obtain past data at the same time,and then displays it intuitively in the form of charts.