Research On Art Style Image Transfer Based On Generative Adversarial Network

With the continuous development of technology and economy,people’s require ments for art experience are also constantly changing.The emergence of image style transfer technology provides a solution to people’s specific artistic needs.The contin uous progress of deep learning makes the technique of image style transfer rise to a n ew height.From the method based on convolutional neural network to the method ba sed on generative adversarial network,the artistic visual effect of image style transfe r is also continuously improved.However,the traditional algorithm model still has s ome problems,and the generated pictures are difficult to meet people’s artistic needs,such as: model training is unstable,the generated pictures lose content information,color distortion,texture contour is not clear,etc.Based on these problems,based on t he framework of the traditional generation adversarial network model,this paper con ducts relevant research and proposes improvement and optimization.The main resea rch contents are as follows:(1)Two datasets of artistic style image transfer are established.We first create a dataset of images of Van Gogh,Monet,Cezanne,Ukiyoe,and real world landscapes from wikiart and Flickr.Secondly,for illustration style,a new illustration style imag e dataset is created by merging foreign public datasets and domestic publishing hous e websites.(2)This paper proposes an artistic style image transfer model based on cyclic ad versarial generation network.In this paper,based on the basic framework of Cycle G AN,the batch normalization technology in the original discriminator network of Cyc le GAN is replaced with the gradient normalization technology to solve the problem of unstable model training,which improves the performance of the discriminator and makes the gradient space of the discriminator smoother,thereby improving the stabi lity of model training.In the process of image transfer,the generated images have co lor distortion and unclear texture contours.In order to solve the problems,this paper changes the original CycleGAN generator network,and introduces a new Res Ne St re sidual network to replace the previous Res Net residual network.Ablation experiment al results show that the gradient normalization and the new residual network are bett er than the previous structure in performance.The comparison experimental results s how that the PSNR value and SSIM value of the generated image are higher than si milar comparison methods,which proves the effectiveness of the proposed model.(3)This paper proposes an illustration style image transfer model combining att ention mechanism.Based on the above improved model,this paper introduces a selfattention mechanism on the generator network to obtain more image detail features a nd improve the quality of the generator for the particularity of the illustration style.Ablation experiment results prove the reliability of the self-attention mechanism,and model comparison experiments also verify the effectiveness of the proposed model.