Recognition Of Cucumber Downy Mildew Based On Deep Learning And Development Of Apparatus

Cucumber has been infected by disease more frequently due to its own habits since the21 st century.Downy mildew,a common disease,would seriously affect the growth and yield of cucumber,which would even spread to the whole planting area.Cucumber downy mildew has become an important factor limiting the yield and quality of cucumber.Therefore,the accurate identification of downy mildew is one of the key technologies to control the quality and quantity of cucumber production.The traditional manual diagnosis method is inefficient and expensive,thus being hard to meet the needs of large-scale planting and production.Therefore,how to intelligently identify cucumber downy mildew with acceptable efficiency and cost has become an urgent problem to be solved.With the application of deep learning in various fields,many researches had been carried out in crop disease recognition and other aspects.Automatic recognition of cucumber downy mildew using deep learning algorithm has become a research hotspot and has made promising progress.In this paper,the identification method and device of cucumber downy mildew were studied to provide technical support for the accurate and rapid identification of cucumber downy mildew.Specific research contents and results are as follows:(1)The image data set of cucumber downy mildew was constructed,and the super-parameter batchsize and optimizer were selected through experiments.In this paper,cucumber downy mildew and cucumber healthy leaf images at different periods of disease were selected as research objects,and the image data set of cucumber downy mildew was established through screening.To increase the difference of samples,geometric transformation and color transformation methods were used to enhance the data set,and maximum and minimum normalization method was used to normalize the data set.Then,to select the optimal hyperparameter and optimizer for subsequent tests,five networks such as MobileNet-V2 were used to establish the model by setting different parameters.The optimal hyperparameter batchsize was selected as 32,and the optimal optimizer was Adam optimizer.(2)A downy mildew recognition model based on improved ResNet50 network structure was proposed.This model took ResNet50 as the basic network architecture,and used feature fusion method to improve its layer.Shortcut was used to convert the output feature map of Layer1 into the same shape as the input feature map of Layer3.Meanwhile,the output feature graph of Layer2 was converted into the same shape as the input feature graph of Layer4,and then these feature graphs were fused to obtain the final output feature graph,which was embedded into the SE attention module to form a new model SE-l＿resnet50.Experimental results showed that the model can achieve the recognition accuracy of 94.37% on cucumber downy mildew image data set,which is 6.2% higher than the original ResNet50 network.(3)The identification accuracy of SE-L＿ResNet50 against cucumber downy mildew was optimized.SE-L＿ResNet-50 combined with ResNet-101,SE-ResNet50 and Inception-V3 were used as sub-classifiers for dual,triple and quadruple-model integration tests.Aiming at the problem of large oscillations in the training process of the four-model integration task,a manual weight improvement method was proposed to reduce the weight of the Inception-V3 network and incorporate its weight ratio into the SE-L＿ResNet-50 model.The test results showed that the accuracy of the integrated model based on the Integration of the four models could reach 95.85%,and the vibration range of integration Model-1 was reduced after the improved weight,and the accuracy could reach up 95.97%,which could realize the accurate identification of cucumber downy mildew disease.(4)Aiming at the problem that SE-L＿ResNet50 model is too large to deploy mobile terminal,a new attention-mechanism module CBAM-Ⅱ was proposed.The spatial attention module and channel attention module in CBAM were changed from serial connection to parallel connection and added together.The structure of the convolutional attention module CBAM was optimized,and the validity and universality of the module in the convolutional neural network model was verified in the cucumber downy mildew image data set and tomato disease image data set respectively.Secondly,the optimal transplantation model was selected by setting different compression ratios of the full connection layer of the Channel attention module.The results showed that the accuracy of MobileNet-V2+CBAM-Ⅱ was the best in cucumber downy mildew data set and tomato data set,reaching 93.32% and 99.47%,respectively.When the compression ratio was 16,the recognition accuracy of MobileNet-V2 in cucumber downy mildew image data set was93.28%,which was closed to the highest recognition accuracy,and the size of the model was only 18.2M.Compared with the original model,the size of the model was reduced significantly,which was more suitable for the deployment of mobile applications.(5)A cucumber downy mildew recognition device based on Raspberry PI was developed.With Raspberry PI as the processing unit,combined with camera,drive motor,storage and Bluetooth modules,the downy mildew intelligent identification car and cucumber downy mildew recognition system were designed.The MobileNet-V2+CBAM-Ⅱ model is embedded in the intelligent identification vehicle and tested in the field.The results showed that the average accuracy of downy mildew recognition system embedded with MobileNet-V2+CBAM-Ⅱ model could reach 93%,which basically met the needs of downy mildew recognition task.