Tea Leaf Diseases Recognition Based On Edge Intelligence

China is a big country in the production,consumption and export of tea.Tea plays a key role in the economy and culture.Tea is also an important cash crops for targeted measures in poverty alleviation in southern mountainous areas of China.At present,there are as many as hundreds of tea tree diseases recorded,of which leaf diseases account for the largest proportion.It is time-consuming and laborious to rely on manual recognition.Image recognition technology relying on deep learning gradually replaces manual identification of crop diseases.Traditional deep learning models are hard to deploy to edge devices owing to their many parameters and high computational complexity.Considering the underdeveloped network in remote mountainous areas,It is crucial to implement efficient monitoring of tea leaf diseases at the edge of the network.This article specializes in eight kinds of tea leaf diseases such as Exobasidium vexans massee,Tea red rust,Pestalotiopsis theae,Tea anthracnose,Tea netted blister blight,Tea red scab,Tea brown blight and Tea algae leaf spot.a lightweight convolution neural network and automatic model pruning(AMC)method are used to train the edge model of tea leaf diseases recognition,which can be deployed on the edge device.The edge inference framework Tensor RT is used to deploy the model on the edge device Jetson Xavier NX by model quantization.The results of this paper can be directly applied to tea garden disease control robots.The main work and results are as follows:1)The establishment of tea leaf diseases identification model based on a lightweight convolutional neural network was studied.A lightweight convolutional neural network,Mobile Net V2,was selected as the network model to be finally deployed on the edge device for the recognition scene of tea leaf diseases at the edge.The method of transfer learning is used to solve the problem of less data sets of tea leaf diseases.The recognition accuracy of the tea leaf diseases detection model after training on the test set can reach 98.58%,and the model size is only 8.74 MB.2)Study the automatic pruning algorithm of the tea leaf diseases identification model.In order to improve the pruning efficiency of the model,the AMC algorithm was used to determine the optimal combination of sparsity of each layer of the model in the pruning process.The accuracy of tea leaf diseases identification was 98.42%,98.19% and 97.83%when the floating-point operations per second(FLOPS)of the model were reduced by 30%,50% and 70%,respectively.The model storage volume is reduced by 41.88%,57.32% and81.69%,respectively.The pruning effect is better than the traditional channel pruning algorithm based on the L1 norm.At the same time,the edge model obtained after pruning can also be used as the transfer learning pre-training model to identify other crop diseases.3)Research on the deployment of tea leaf diseases identification model at the edge.Firstly,the tea leaf diseases identification model was deployed by the combination of edge inference framework Tensor RT and edge equipment Jetson Xavier NX to enhance the inference speed of edge model.Then,the pruned model is deployed with Tensor RT,and it is verified that within a certain range,the higher the pruning rate of the model,the faster the model inference speed will be.Finally,a hybrid acceleration scheme of model pruning and FP16 quantization was used to further improve the model inference speed with little loss of model accuracy.