Research On The Improvement Of Image Ink Style Migration Network ChipGAN

Image style transfer is a research hotspot in the field of computer vision,which is widely used in artistic creation,film and television entertainment.As one of the important representatives of traditional Chinese painting,Chinese ink painting is an important cultural treasure of China.Using computer technology to automatically generate Chinese ink painting can not only broaden the application field of computer,enrich the creation way of ink painting,but also have far-reaching significance for carrying forward Chinese traditional culture.Because style transfer has a wide range of application scenarios,more and more style transfer models are proposed.However,the ink style images generated by these models have some problems,such as low image quality and poor white space.Therefore,based on the Chip GAN model,this thesis improves the image detail information extraction and white space effect,so as to improve the quality of ink painting style images.The specific research content of this thesis is as follows:(1)By replacing Res Net residual network with residual-dense network RDN in generator,Patch GAN discriminator is changed into multi-scale discriminator to solve the problem of poor image quality generated by Chip GAN’s image style migration.The residual-dense network RDN is used in the generator so that the network can reuse the shallow features of the image continuously.By combining the shallow feature and deep feature,the feature image extracted by the generator contains more abundant feature information.The discriminator is changed from the original Patch GAN to multi-scale discriminator.Multi-scale discriminator combines multiple Patch Gans to enable the network to obtain the discriminant results of images at different scales,thus improving the discriminant ability of the whole discriminant network.The experimental results show that the improved model can generate ink style images well,and the images have high image quality.(2)In order to solve the problem of generating unnatural white space effect,the white space loss is added on the basis of the original loss function.In order to prevent the generated image from producing mottled color blocks in the blank area,this thesis conducts threshold segmentation on the original image and the generated image to segment the background white space in the image.In addition,L1 loss and SSIM loss are introduced to restrict the background white space,so that the network can better generate the ink image with white space.Experiments show that the image generated by the model is more consistent with the style of ink painting by constraining the model with the increase of white space loss.At the same time,a ink style transfer system is designed to realize the ink style transfer of any image.