Research And Implement Of Image Dataset Expansion Based On Adversarial Network

In recent years,deep learning has been a hotspot in the field of computer.With the deepening of research,deep learning has reached the bottleneck period in theory.The quality and size of the data set are critical to the impact of the final performance of the deep learning model.The generative adversarial networks(GAN)is a new unsupervised generative model,which trains neural networks with its novel game theory,and it has been widely concerned by researchers since it was proposed.There is a generator network and a discriminator network in GAN,which can learn from each other through the continuous confrontation between the generator and the discriminator until the Nash equilibrium is reached,so that the image data can be generated from scratch,so the image data set can be expanded by relying on the powerful generation ability of GAN.However,in the experiment,GAN has some problems,such as unstable training,difficult convergence of the model,high freedom and uncontrollability of the training process,which makes the quality of the generated image often unsatisfactory.Deep Convolutional GAN(DCGAN)is a derivative model of GAN.Based on GAN,the deep convolutional neural network is used to improve the network structure of GAN,which makes the training process more stable and becomes the standard structure of various variants of GAN.In the experiment,when DCGAN generates 64×64 resolution image samples,some meaningless images will appear,or some features of the images are similar and lack of diversity.In DCGAN,the cross entropy method is used as the loss function,and in theory,when the least square method is used as the loss function,which can punish the forged samples that are judged to be true by the discriminator,but it is far away from the decision boundary of the discriminator,so that they are gradually close to the decision boundary,in this process,the real data distribution is constantly fitted to improve the quality of the generated image.Therefore,starting from the loss function,this thesis uses least square as the loss function of the network model to improve DCGAN,and design a deep convolution generative adversarial networks based on least square—LS-DCGAN.The comparative analysis of the experimental results of DCGAN and LS-DCGAN proves the feasibility of the improvementBecause the high-resolution image has more complex spatial characteristics,it is difficult for the generator to fit its data distribution,at present,GAN is less effective in generating high-resolution images larger than 64×64,and it is easy to produce checkerboard artifacts This thesis introduces residual block on the basis of LS-DCGAN to properly deepen the network depth,which helps to improve the learning ability of the network,and design the size of the transposed convolution kernel so that it can divide the step size in order to reduce the effect of the checkerboard artifacts on the generated images,thus design a deep convolution generative adversarial network based on residual block based—Res-DCGAN.Using LS-DCGAN and Res-DCGAN to generate 128 × 128 and 256 × 256 high-resolution images for comparative experiments,it is proved that Res-DCGAN has more advantages in generating high-resolution images.