Automatic Identification Of Crop Image Types Based On Remote Monitoring Points

With the advancement of the "Internet+" agricultural project,a large number of fixed-point cameras are set up to collect crop image information in the agricultural park,and a large number of untagged crop images were produced in the process.So far the number of images has reached a level that can not be recognized by manual visual recognition.So there is an urgent need for an effective method to realize automatic identification of crop image types.In the past crop recognition methods such as support vector machine(SVM)and so on were used to extract artificial features from samples obtained by remote sensing and other techniques.According to the related research,support vector machine is one of the better classifiers,but it is suitable for small samples;Manual observation needs a large number of labor force,recognition accuracy needs to be improved;Remote sensing collection range is wide but vulnerable to environmental impact.With the development of image recognition technology,the method of machine learning for crop image recognition has been gradually applied.Among them,convolution neural network is one of the most popular machine learning algorithms at present,and it is widely used in image recognition and other fields.Therefore,this paper uses the convolution neural network algorithm is used to realize crop species image recognition.First of all,based on the monitoring point camera to collect soybean,potato,rice and corn crop images,the data set was established according to the visual classification of crop morphology in each growth period,the data set is randomly divided by hierarchical and isometric sampling method.Among them,the training set of small sample data set is 400,and the test set is 200.The training set of large sample data set combined with image augmentation technology was 12620,and the test set was 6310.In order to save the network running time,the segmented images are preprocessed,including random cutting,scaling,grayscale and so on.Secondly,this study established three crop image recognition models,including: model migration learning based on Alex Net deep network model;Combined Alex Net depth model and PSO algorithm;According to the structure of lenet-5 model,a crop identification network model(lenet-m)suitable for this study was designed.In order to test the generalization ability of the model,this study used small sample data set for experiment,and referred to relevant studies.In order to prevent model overfitting,Image Net dataset is used for model training in advance.Referring to the related research,the equivalent experiments of five iterations are carried out for the three models.Because Le Net-m is a lightweight model,the number of iterations of the model is further set to 50 and 100.Thirdly,the optimal recognition results of Alex Net model migration learning,Alex Net and PSO(particle swarm optimization)and Le Net-m are 85.94%,92.97% and 95.5% respectively.Based on the recognition rate of the three models,taking into account the model running time,hardware requirements,model performance and other aspects,this study optimized the model parameters of Le Net-m,including batch size,momentum factor and sample data volume.The experiment shows that the recognition result of Le Net-m model based on big data set can reach99.38%.Finally,in order to enhance the usability of the model and realize the automation of crop type image recognition,this study designed the crop image recognition system based on Matlab GUI.It is of certain research significance to realize remote monitoring and intelligent decision management for the growth and health status of crop species image targeted in the later stage.