Zero-Shot Ink Wash Painting Generation Based On Deep Learning

Chinese traditional painting is an ancient and unique form of artistic expression.Compared with Western painting art,Chinese traditional painting art pays more attention to the visual effect of charm,makes good use of lines,and pays less attention to texture and other information.As one of the traditional Chinese painting types,ink wash painting can be roughly divided into figures,flowers and birds,animals and landscape painting according to its subject matter and objects of expression.It depicts the shape,light and shade of objects in an expressive way,both realistic and freehand,which can be summarized as the two characteristics of "GongBi"and "XieYi" Through the use of ink,artists create light and shade matching,alternating density and diverse colors,making ink wash painting a rare art treasure in the world.Image style transfer task is a very active research direction in the field of computer vision in recent years.Because of its diversity of transfer results and the idea of feature extraction,it has certain academic and application value in computer vision applications.In recent years,more and more researchers begin to pay attention to Chinese traditional art,trying to use computer vision and graphics technology to protect and publicize these traditional culture and art.At the current stage,most of the leading neural style transfer research tasks are combined with Western painting art works,which generate higher quality results in the content preservation and style transfer parts.Recently,there have been some style transfer methods to apply traditional Chinese painting styles(such as ink wash painting styles)to real photographs.However,there are some limitations to using these style transfer methods to ink wash stylize real photos of different object types in the dataset For example,when the input content image is a type that the generator has not seen before in the training stage(i.e.zero-shot data),the generated image will retain the content features of the data samples in the training set,resulting in distortion and deformation of the generated image.Therefore,this paper attempts to decouple style and content from feature representation,and proposes a framework for ink wash style transfer based on GAN,named Style-woven Attention Network(SWAN).The framework contains a style-woven attention network,which reduces the influence of style images on the semantic content images by decoupling the content features and style features.Through the self-attention mechanism module,style features are woven and semantic features of content images are enhanced.The ink wash painting style is summarized into three characteristics and realized by deep learning.Finally,the task of ink wash painting style transfer with zero-shot is realized.The model generates high-quality ink wash painting style images.In the training stage,the model decomposes the representation of data samples in an unsupervised way to reduce the semantic correlation between content and style.In addition,the ink wash style loss can improve the style encoder’s learning ability of ink wash style.Finally,in order to verify the stylization capability of the method proposed in this paper,ChipPhi,an existing public dataset for the ink wash style transfer task,is modified and supplemented.Based on a large number of experiments,the method proposed in this paper performs better than other style transfer methods on visual quality,three quantitative evaluation indicators(FID,Kernel MMD and SSIM),user study and generalization experiment,which achieved good results in content preservation and ink wash style learning and transfer.