Research On Crop Disease Image Data Augmentation Method Based On Generative Adversarial Network

With the development of deep learning and computer vision technology in recent years,the method of using crop leaf images for disease recognition has been widely studied.The convolutional neural network model for crop disease recognition requires sufficient data to achieve better results.However,the leaf data of crop diseases is difficult to obtain.At this time,the crop disease recognition model based on deep learning and convolutional neural network is very susceptible to stuck in overfitting.Traditional data augmentation is a method of expanding the training set size by performing geometric affine transformation or color space distribution transformation on the basis of original data.Data augmentation can improve the generalization ability of the model and improve the performance of the model in actual application scenarios.However,the traditional data augmentation method is relatively simple.Although it can improve the performance of the model to a certain extent,simple transformation cannot enrich the distribution of the data in the high-dimensional feature space.In recent years,generative adversarial networks have shined in the field of image generation and image-to-image translation.This thesis attempts to use an unsupervised image translation method based on generative adversarial networks to convert crop healthy leaf images into diseased leaf images,as a data augmentation method in the task of few shot crop disease recognition.The main work of this thesis is as follows:(1)First,this thesis proposes an unsupervised image-to-image translation model based on CycleGAN,which is guided by the class activation maps.An auxiliary classifier is introduced into the generator structure of CycleGAN.The auxiliary classifier uses the generator’s encoder feature map for distinguishing healthy leaf images from diseased leaf images.The auxiliary classifier will provide the weight parameters for generating class activation maps.Similarly,the same auxiliary classifier is introduced into the discriminator.This thesis constructs an image translation dataset of crop leaf images containing 6 different varieties based on PlantVillage,and conduct image translation experiments on this data set.Experiments verify the effectiveness of this model.(2)In the improved model of(1),this thesis proposes a reconstruction loss function based on the feature map of the encoder module of generator.This thesis verified its effectiveness in the task of few shot leaf image translation task,and successfully used the redundant information in healthy leaves to generate a large number of diseased leaf images.Finally,by comparing with traditional data augmentation methods,the experimental results show that the model in this paper can be used as a data augmentation method to expand the crop disease recognition data set and effectively improve the recognition accuracy of the convolutional neural network.