Research On Plant Leaf Image Disease Recognition Method Based On Neural Architecture Search

At present,the world are facing a serious food crisis.A timely and accurate grasp of the types,severity and development of plant diseases can effectively reduce the economic losses caused by diseases to agroforestry production.The plant disease recognition method based on traditional computer vision and image processing requires complex image preprocessing and feature engineering,and its practicability and recognition effect are not ideal.The recognition method of plant leaf diseases based on deep learning has the advantages of high accuracy and fast speed,and provides effective technical means for accurate and rapid analysis of plant diseases.However,the existing deep learning model for plant disease identification mainly relies on a general artificially designed network architecture,which is difficult to meet the application requirements for rapid and accurate identification of plant disease.In this study,a disease recognition method from plant leaf image was proposed based on the Neural Architecture Search(NAS).The main research contents of this paper are as follows:(1)In this paper,we study the method of plant disease recognition based on neural architecture search.Firstly,the principle of the neural architecture search method are analyzed,including search space,search strategy and performance evaluation strategy.Then a neural architecture search method based on Bayesian optimization guided network morphism is constructed.Finally,a total of 54306 plant disease images including 14 crops and 26 diseases of PlantVillage were used as experimental data.The generalization ability of the model was tested with a data set of maize disease leaves containing a real scene and a fixed background.The experimental results verify the feasibility and effectiveness of the proposed method.(2)The influence of unbalanced data and lack of color information on the recognition results was analyzed by balance processing and gray scale transformation of training data.Different search times and initial layers of neural network architecture were set in the process of neural architecture search to study the effect of neural architecture search model on plant disease recognition.Tomato disease data set and maize disease data set were used for experiments.The results show that the proposed method can search the network architecture with good performance in a relatively short time.The balance of the training sample data has little effect on the recognition result,the color information of sample data is helpful to improve the accuracy of plant disease identification,and the number of convolutional layers of the initial network architecture has almost no effect on the optimal network architecture generated by the neural architecture search.(3)In order to solve the problem of fine-grained plant disease image recognition,the LeakyRelu function was used as the activation layer for neural architecture search to improve the methods.At the same time,data enhancement techniques including random cropping and flipping were used to expand the tomato disease images and apple disease images respectively,compare and analyze the experimental results.The results show that the improved measures adopted in this paper can improve the recognition accuracy of fine-grained disease images.