Research On Agricultural Pest Identification Based On The Deep Learning Model

In the world,many agricultural problems are caused by crop pests.Improper pest control is the main reason for the increase in pest resistance and the excessive pesticide residues in agricultural products.Accurate recognition of pests is the key to pest control and is a considerable challenge in farming.Deep learning models have shown great promise in image recognition,drawing the attention of many agricultural experts.However,the lack of pest image datasets and the inexplicability of deep learning models have hindered the development of deep learning models in the field of pest recognition.Therefore,this paper first constructs the pest image data set,and uses the migration learning and fine-tuning training depth learning model to improve the accuracy of model recognition.Secondly,a negative sample test method is proposed to recover the uploaded pest image data and expand the pest image database.Finally,the visual analysis of the model is carried out to explore the interpretability of the deep learning model.The research results are as follows:(1)An effective pest image data set is constructed for agricultural pest image recognition.The data set,named IP67,contains 67 types of agricultural pests and 67953 original agricultural pest image data.(2)This paper selects four models: Inception-V3,Res Net-50-V2,Mobile Net-V2,and Xception,which have made remarkable achievements in the field of image recognition.Using the method of combining transfer learning and fine-tuning training,the depth learning model is trained based on the constructed pest image data IP67 and the currently known public data set ip102 with the largest number of pest species,and the effect of the depth learning model on the pest image data set is verified.By comparing the recognition results of IP67 and ip102,it can be found that the new data set is more effective.(3)To expand the pest image database,a negative sample test method is proposed in this paper.Through the threshold setting,the model can identify positive and negative samples and selects the most suitable model to collect the pest image data uploaded by users.The experimental results show that Xception is the most suitable model for negative sample screening among the four deep learning models,and the optimal threshold is 0.55,which can reach the positive sample retention rate of 83.13% and the negative sample exclusion rate of 81.17%.(4)Taking the deep learning model inception-v3 as an example,this paper visualizes its internal hidden layer and decision-making layer by combining the methods of feature analysis,model inspection,salient representation and activation mapping.The experimental results show that:(1)The convolution layer changes from shallow to deep,and the concept type learned by the convolution neural network from pest image also changes from low level to high level.The shallow neural network learns simple general features such as texture,color,and edge,which are easily learned by the neural network from all kinds of pictures.(2)Deep neural network is to learn visual semantic features,that is,the matching features of image and correct label.Such characteristics need specific training and learning for specific tasks.(3)A reasonable explanation for the recognition decision of the deep learning model is given.Through the confusion matrix obtained by visualizing the decision-making level of the model and combining the model recognition results,it is found that the model pays more attention to the local characteristics of pests in pest image recognition,which is a recognition process from local to overall.In conclusion,this paper constructs a new agricultural pest image data set for intelligent pest recognition.Four typical deep learning models are trained on this data set,and the negative sample test is carried out.It is concluded that among the four deep learning models,Inception-v3 has the best recognition effect,and Xception is the most suitable model for negative sample screening.At present,the most important thing is to collect data and train an intelligent model that can guide users to prevent and control pests at present.Therefore,Xception is recommended as the embedded mobile terminal model,and Inception-v3 is recommended as the basic model for the interpretability research of the deep learning model.Based on of previous studies,this paper further expands the pest image data and studies the deep learning model.These results will contribute to the research in the field of agricultural intelligent pest recognition in the future.