Research On Rice Disease Recognition Method Based On Convolutional Neural Network And Ensemble Learning

Breeding informatization can significantly improve breeding efficiency and management level,and is an important direction for the development of modern seed industry in China.Accurate identification of crop disease types is the key to achieving information technology for disease resistance identification in variety resistance identification experiments.In recent years,deep learning technology has achieved great success in image classification applications,providing effective techniques for accurate recognition of crop disease images.However,existing classification models still face problems such as complex network structures,difficulty in training network models,and the need for huge hardware computing resources support.In response to the aforementioned technical issues and practical application requirements,this paper takes typical rice diseases as an example to conduct research on rice disease identification methods based on convolutional neural networks and ensemble learning,providing new technologies for accurate identification of crop disease resistance.The main research content and innovation points include:(1)To understand how deep structured network models represent disease images,a convolutional layer feature visualization analysis method based on deep convolutional neural networks is proposed.Firstly,taking an 8-layer pre trained lightweight Alex Net network as an example,an automatic rice disease classifier is constructed by automatically learning image features from raw disease image data;Then,using the improved deconvolution feature visualization method,quantitatively analyze the characteristics of each intermediate layer of the Alex Net network,and explore the specific process and mechanism of Alex Net network characterizing diseases.The experimental results show that convolutional neural networks can effectively and automatically extract disease image features.The research results can lay the foundation for the selection and design of subsequent convolutional neural network models.The research results can lay the foundation for the selection and design of subsequent convolutional neural network models.(2)To achieve efficient and accurate recognition of rice disease images,a lightweight rice disease recognition method based on attention mechanism and Efficient Net,CGEfficient Net,is proposed.Firstly,a lightweight convolutional attention module is introduced to improve the main module of Efficientnet-B0,which flips the bottleneck convolutional kernel with lightweight,maintaining high accuracy while making the model lighter;Then,use the Ghost module to optimize the convolutional layer in the network,reducing the number of parameters and computational complexity of the network;Finally,use the Adam optimization algorithm to improve the convergence speed of the network.The results showed that for the dataset of typical rice disease images,the average accuracy of this model was 95.63%,which is superior to other classic classification models.(3)To further improve the accuracy and generalization of disease identification models,a rice disease identification method based on improved Stacking ensemble learning is proposed.Firstly,the proposed lightweight recognition method CG-Efficient Net is used as a single classifier;Then,based on the perturbation of training samples and algorithm parameters,the single classifier is trained to increase the differences between the single classifiers;Finally,the weights of the base model are learned by minimizing the loss function of SLSQP,and the samples are classified by Stacking weighted integration.The experimental results show that compared with a single classifier,the weighted ensemble based on SLSQP has a higher recognition accuracy of 96.1%,effectively improving the classification accuracy of weak classifiers and performing outstandingly in solving the problem of rice disease recognition.The method proposed in the paper has improved the accuracy and generalization ability of the model to a certain extent.