Image Sample Enhancement And Generative Study Of Leaf Diseases And Pests Based On Generative Adversarial Network

Due to the frequent occurrence of pests and diseases,crops in our country are facing serious threats.The types of pests and diseases are diverse,the severity is high,the speed of transmission is fast,and the coverage is wide.Inaccurate or delayed diagnosis of pests and diseases can lead to misuse and abuse of pesticides,which not only reduces production but also seriously pollutes the environment.Therefore,it is very important to realize timely and accurate diagnosis and control of pests and diseases.Deep learning provides efficient techniques for identifying agricultural pests and diseases.However,deep learning model training requires a large amount of image data of pests and diseases,and the collection of agricultural pests and diseases data is difficult,and many samples of pests and diseases are scarce,which is one of the challenges of agricultural pests and diseases controlIn view of the high cost of obtaining pest and disease image data,image generation technology has become increasingly important.In the image generation task,the goal is to obtain images with corresponding semantic content after training the generation model.Currently,visual or image processing research,such as image rain removal,image fog removal,image conversion,and image super-resolution reconstruction,has applied the principle of image generation.However,due to the nonlinear relationship between complex images and the inherent generalization bias of regular convolutional neural networks,existing methods are difficult to accurately capture the pathological information required to generate important pest and disease images.To address this challenge,the main work of this paper is as follows:Firstly,many current generation networks are constructed by convolutional neural networks.However,the spatial operations of CNN have a local nature,resulting in low efficiency in capturing the correlation between long-range features.It is difficult to generate complex geometric object images completely and the generation effect is poor.To address these issues,a generative adversarial network(GAN)with fusion attention mechanism was proposed to enhance images.The method added an attention mechanism to improve the distant features of images and connect these features to generate coordinated objects,thereby stabilizing the training of GAN.The attention mechanism includes spatial attention and channel attention.By combining the spatial attention and channel attention mechanism,the correlation between features can be learned more effectively,partially solving the problem of CNN’s inability to capture the correlation between long-range features.It can effectively improve the enhancement effect of images.To verify the effectiveness of the fusion attention mechanism for image enhancement,the designed attention mechanism was added to the SRGAN and Deblur GAN,where SRGAN performs image super-resolution reconstruction and Deblur GAN performs image de-blurring.The experiment results showed that the proposed method further improved the stability of GAN training and achieved good results in terms of image enhancement,thus verifying the feasibility of the method in application.Secondly,in order to address the global feature correlation problem of pest and disease images in different environments and achieve data expansion and class balance,this study proposed an image generation model based on DCGAN.The model uses the DCGAN generator to learn the data distribution of different pest and disease types,thereby generating new pest and disease data.Compared to the problem of global feature absence caused by the generalization bias of regular convolutional neural networks,this study uses a deep convolutional neural network to build a generative adversarial network,and the discriminator module constraints the generator to pay attention to both global and local structural information.The experimental results showed that the method can consider the global and local features of pest and disease images while generating high-quality pest and disease images,and achieved good results on mainstream pest and disease image datasets.