| Agriculture is the basis of human survival,and rice,as the main economic crop in China,plays a vital role in agricultural production.With some climatic and environmental problems,rice is easily affected and invaded by diseases and insect pests in the process of growth and development,incidence of a disease is getting higher and higher,especially the pest disasters caused interfere with the growth of rice,which makes the quality of rice and yield has been seriously affected,so the control of rice pests is extremely important.The traditional identification method of rice pests in China has a long cycle and low identification rate,which is not conducive to the timely prevention and control of rice pests.With the continuous progress of artificial intelligence technology,image processing,big data and deep learning technologies have been widely applied in the agricultural field.The combination of information technology and agriculture provides a new ideas for rice pest identification.It is difficult to collect rice pest image dataset,and it needs to be augmented to expand the quantity.Traditional data enhancement methods include translation,clipping,and flipping,etc.The images generated by these methods are difficult to meet the requirements of rice pest image dataset.With the introduction of generative adversarial network,it provides a new research direction for data enhancement.In this paper,the image of rice pests is taken as the research object,and the improved generative adversarial network is used to expand the data.On this basis,deep learning technology is used to identify rice pests.This study is of great significance to the improvement of rice yield and national food security.The main research contents and achievements are as follows:(1)When the original GAN network is used to generate images of rice pests,there will be many problems such as stability in the training process and generation quality.To solve these problems,the Capsule GAN network is used to generate images of rice pests,and the CNN of the discriminator in the Capsule GAN network is replaced by Caps Net,which can ensure stable and high-quality images.Capsule GAN networks use Caps Net as the discriminator network.Through its unique dynamic routing algorithm between capsules,the discriminator can learn the rich representation of image features and their relationships in the global scope during inverse image coding.However,the discriminator ignores important local features and larger receptive fields.Therefore,Pyramid Split Attention is added to improve Capsule GAN network,which is used to extract multi-scale spatial information effectively and generate long-term channel dependence.Compared with generative adversarial network,the improved Capsule GAN network generated better image quality and higher definition of rice pests.(2)Aiming at the problem of low identification accuracy of rice pests,a rice pest identification method based on the improved residual network model was proposed based on the improved Capsule GAN network for the data enhancement of rice pests.In this model,dynamic routing capsule structure is embedded into the deep convolution model of the residual network to replace the full connection layer of the residual network.Firstly,feature maps are extracted through four residual blocks,and the feature maps are encapsulated and encoded.Secondly,inter-layer routing is carried out to reduce the large amount of information lost in the output of CNN.Through the identification of 14 kinds of rice pests,the effects of different parameters,such as learning rate,batch size,activation function and optimization combination,were analyzed.The experimental results showed that the accuracy of the improved residual network model reached 77.12%,and the average detection time of a single plate was0.32 s.The improved model meets the requirements of rice pest image recognition,effectively improves the recognition accuracy,and provides a feasible scheme for rice pest identification in the actual agricultural scene. |
PDF Full Text Request