| Rice is the main food crop in China,the production of rice is involved in the people’s food and clothing problem,but also involved in the development of national economic level.In recent years,due to the weak resistance of rice plants to insect and disease resistance,as well as the deterioration of farmland ecological environment caused by economic and social development,the rice production has been increasingly influenced by diseases and insect pests,which has also brought great economic losses to rice growers and the country.Traditional artificial pest identification methods have low efficiency and poor real-time performance.Traditional machine learning relies heavily on feature extraction and has poor universality.In recent years,deep learning technology has been widely used in the field of image recognition,which can realize crop pest identification in an end-to-end manner,overcoming the shortcomings of traditional identification methods such as low efficiency,poor realtime performance and cumbersome feature extraction operations,but the accuracy and robustness of recognition still need to be improved.In order to find a suitable identification method for rice pests,this paper carried out identification research on rice pests based on convolutional neural network in deep learning.The main work is as following:(1)In order to accurately identify rice pests,an improved rice pest identification model based on Alexnet was proposed.Alexnet model was improved based on image data sets of four rice pests.The improvement methods include removing the original LRN layer on the basis of Alexnet model,adding a batch normalization layer after the convolution layer to alleviate gradient dispersion and make the model more stable,replacing the original full connection layer with global leveling,reducing model parameters and reducing the possibility of over-fitting of neural network model training.PRelu function was used to replace Relu function to avoid neuron necrosis.The experimental results show that: 1)The recognition rate of the improved model on the pest data set is no less than 98%,which is 1.96% higher than that of the original network and higher than other traditional networks,such as Le Net5,VGG13,VGG16.2)The loss value of the improved model is stable around 0.03,0.1 lower than that of the original network,both lower than Le Net5,VGG13,VGG16 and other traditional networks.The experimental results show that the improved method has higher recognition rate and better robustness in rice pest classification.(2)Transfer learning was introduced to solve the problem of less rice pest image data.A rice pest identification method based on convolutional neural network and transfer learning was proposed.In this method,VGG16,which is more suitable for transfer learning,is selected as the benchmark model.Firstly,it is pre-trained on Image Net data set,and then the global average pooling mentioned in the previous experiment is applied to the optimization and improvement of the top layer of VGG16.Then,the improved VGG16 model is applied to the recognition of image data sets of six rice pests.The weight parameters obtained by pre-training are used to initialize the parameters of the improved network before training.During the research process,a large number of comparative experiments were also carried out.Through comparison experiments,it was found that after introducing transfer learning,the top layer of the VGG16 model was designed as a global average pooling layer followed by an output layer,and part of the convolutional layer was unfrozen for training.This way to get the best training effect,with recognition speed,high accuracy,portability and other advantages.In general,this paper preprocessed the rice pest data set by taking images of rice pests acquired by Internet search engine and camera as samples,and proposed two pest identification methods.The improved model based on Alexnet has achieved some results.However,because of the small number of Alexnet network layers and limited feature extraction capability,a deeper model may be used in practical application.The deeper the model layers are,the more data sets are required.Therefore,a method combining transfer learning with convolutional neural network is proposed to solve the problem of insufficient training caused by small samples.Both methods have achieved good identification effect and are universal,providing effective reference for the identification and classification of crop pests. |
