Research On Image-to-ink Painting Translation Based On Generative Adversarial Networks

Style transfer is a very active research direction in recent years.Because of the diversity of transfer results and the generality of the feature extraction ideas in other computer vision tasks,this research has demonstrated certain application value and academic value.As an important cultural heritage in China,ink painting has a strong artistic style and is an important representative of spreading Chinese culture.Research on the integration of technology and art has been an academic hotspot in recent years.In order to better promote national culture,we combines style transfer with Chinese ink art painting,and proposes a new ink style transfer algorithm based on Cycle GAN(Cycle GAN).Experiments show that this method can transfer photos to ink style well by studying the typical features of ink painting,such as stroke,blank and ink diffusion.The contributions of this article are summarized as follows:First of all,aiming at the problems of incomplete semantic information,partial color block errors,and uneven ink transition in the image generation of existing migration methods,a single cycle network structure combined with Adaptive Instance Normalization(Ada IN)is designed.In this structure,the adversarial loss and cycle-consistency loss are combined to complete the image semantic information.This is because the adversarial loss can learn the approximate distribution of real data,and the consistency loss can reduce the mapping path between data domains;The abstraction of the blanks in ink paintings indicates that the model should learn lower-level features,and even discard some of the features to leave blanks.However,the two-cycle structure in the baseline model has strong constraints that lead to the model learns higher-level features.Therefore,the one cycle structure is used to alleviate this constraint to produce natural blanks;In order to achieve the diffusion effect of ink molecules,this article introduces the Ada IN module after the encoding process of the generative network,which integrates the encoded feature maps of the content image and the style image so as to introduce the style information of ink painting during the decoding process.Secondly,in order to further improve the quality of generated ink images,an optimization objective combining Multi-Scale Structural Similarity(MS-SSIM)loss is designed.The combination of MS-SSIM loss and L₁loss in the reconstruction part can not only strengthen the constraints on the reconstructed image from brightness,structure and contrast,but also compensate for the brightness change and color deviation to achieve high-quality image generation.Finally,in order to verify the feasibility and superiority of proposed method,this article compares with the existing Distance GAN,Cycle GAN,Chip GAN models on serveral evaluation indicators such as FID,Kernel-MMD,PSNR and SSIM,the experiment results show that the method in this paper can well complete the task of style transfer from real photos to ink paintings,and generate images with ink style characteristics well.