| In recent years,generative AI has gradually become a research hotspot.In the field of image generation,both methods based on Generative Adversarial Networks and methods based on diffusion models can generate realistic images.Chinese landscape painting holds an important position in Chinese culture and has profound artistic value and historical and cultural significance.Image generation has a wide range of applications in the art field,such as art style transfer.However,there is little work on the style transfer of landscape paintings in the field,and the landscape paintings generated by applying existing transfer models have a certain gap with real landscape paintings in terms of artistic conception,charm,and freehand style.This is because these models rarely pay attention to the aesthetic characteristics of landscape paintings,such as the blank space,color,and composition.In response to the above problems,this thesis innovatively proposes a style transfer framework dedicated to landscape paintings,which includes two steps:The first step is to transfer the contour of the scenery image to the contour of the mountains and rocks.The objects in the scenery pictures are mostly realistic,while the landscape paintings are freehand.Therefore,the landscape paintings generated directly from the pictures are more rigid.To make the objects in the generated landscape paintings have a freehand style,this thesis proposes the EdgeGAN model for freehand style transfer of mountain and rock contours-first,extract the mountain and rock contours in the landscape paintings,design an algorithm to simulate ink diffusion,and obtain mountain and rock contours with a freehand style.Then,based on the improved CycleGAN framework,the contour of the scenery picture is converted into the contour of the mountain and rock.Experiments show that the generated mountain and rock contours have a higher richness of brushstrokes,and the landscape paintings generated based on this contour remove the influence of the object contours in the original picture and have a more distinct freehand style.The second step is to transfer the mountain and rock contours to landscape paintings.The blank space,color,and composition in landscape paintings are usually different from other painting forms.To make the generated landscape paintings conform to these characteristics,this thesis proposes the PaintGAN model for landscape painting style transfer under quantified data supervision-extracting statistical laws such as blank space ratio,color distribution,and composition rules from famous paintings,converting these laws into high-dimensional features of images,and using these features to supervise the generation process.Experiments show that the paintings generated using the PaintGAN model perform better on evaluation indicators such as FID.Meanwhile,this thesis conducts an aesthetic evaluation of the generated landscape paintings,creates a multi-dimensional evaluation dataset for landscape paintings,and establishes an aesthetic quality assessment model for landscape paintings.Experiments show that the landscape paintings generated in this study have a certain aesthetic value.In addition,in order to facilitate the implementation of this research,the two style transfer models are combined,with the backend using the Flask framework and the frontend using the Vue framework,to create a landscape painting style transfer system for aesthetic researchers and landscape painting enthusiasts.After testing,the system is fully functional,easy to use,and has high security and reliability.The two-month trial operation shows that the system has good user experience and complete functions. |
