Research On Crop Disease Image Data Augmentation Based On Generative Adversarial Networks

Deep learning models usually need a lot of data to train for achieving better performance.However,in the field of smart agriculture,it is often very difficult to obtain a large number of high-quality crop disease samples,limiting the development of deep learning in the field of agriculture.Although traditional image enhancement techniques are often used for data augmentation,the data obtained by this method are often not diverse enough.In recent years,generative adversarial networks(GANs)provide a novel and efficient approach for image enhancement,which can generate images with diversity and authenticity,thus facilitating downstream modeling tasks.In this thesis,two crop disease image enhancement problems are studied and two image enhancement algorithms are proposed.The main work is as follows:(1)To solve the problem of unstable training of GAN on crop disease data,a GAN data enhancement model SGCD based on Convolutional Block Attention Module(CBAM)and Drop Block regularization is proposed.The main idea of the model is to use the CBAM attention mechanism to help the generator improve the ability of data fitting,and transfer important information to help the generator make better use of the feedback of the discriminator to improve the training stability of the GAN model.And the Leaky Re Lu activation function is used to replace the Sigmoid activation function in CBAM to prevent the model from gradient disappearance.Meanwhile,Drop Block regularization is introduced to control the discriminative ability of the discriminator.The stability of the model is evaluated by the amplitude of the loss function and the overall performance of the model,and the results show that the SGCD model is more stable on a variety of crop disease data.(2)Aiming at the problem that it is difficult for GAN to generate leaf details on crop disease data sets,a GAN data enhancement model FHD based on double discriminators is proposed.The model is based on the Frequency Domain Image Translation(FDIT)framework to design a double discriminator.The first discriminator is called the original discriminator,which is used to distinguish the authenticity of the real image and the generated image.The other discriminator is the high-frequency discriminator,which first needs to extract the high-frequency components from the real image and the generated image,and then send them to the high-frequency discriminator to distinguish the authenticity of the high-frequency components in the real and generated images,so as to improve the ability of the generator to generate high-frequency details.In addition,in order to make the generated image more realistic,the high frequency discriminator uses the least squares loss function.Experimental results show that the model can generate more image details on a variety of crop disease data,and the image quality is higher.Finally,the data enhanced by the traditional data enhancement method,the FDIT model and the FHD model are verified on the disease classification task,and the results show that the data enhanced by the FHD model can improve the classification accuracy of the model more effectively.