Chinese Painting Landscape Style Transfer Based On Deep Convolutional Neural Network

With the development of deep learning network,image style migration has become one of the hottest directions in the field of ARTIFICIAL intelligence.As an effective combination of computer science and art,image style transfer algorithm endows computer with the ability of re-creation.Image style transfer technology not only keeps innovating and iterating in the academic field,but also has been applied developed in the industry.It shows great potential in commercial and artistic aspect.Nowadays,the research of this area mainly focuses on the overall image transfer,especially on the western paintings.however,there are less research on part image and Chinese traditional paintings.As a sub-task of style transfer task,target style transfer has different characteristics and requirements from global style transfer.The target style transfer task could be disassembled into three sub-tasks:target extraction,style transfer and image fusion.The target extraction task means we need to extract the target object from the image,including its classification information,location information and spatial layout information.The style transfer process is similar to that of the global style transfer.The image fusion task requires to embed the generated landscape stylized image into the original image according to the original position information of the target object.Moreover,we use the image fusion algorithm to deal with the embedded area to make the fusion smoother and more natural.Most of the existing researches are carried out independently,the integrated researches are relatively few.In achieving a specific target style transfer task,we need to achieve a balance between speed and effectiveness.Aiming at the technical requirements of each sub-task in the style migration of landscape painting,the research implements the target style transfer algorithm based on Mask R-CNN.The main research work is shown as below:(1)For the classical target detection algorithms,such as R-CNN,Fast R-CNN and Faster R-CNN,systematically studied were taken.Finally,the target style transfer was realized based on the Mask R-CNN algorithm framework.In this paper,the core principle and algorithm flow are introduced,in addition,the Keras and TensorFlow learning frameworks are studied,and the code implementation of Mask R-CNN and its key steps are studied in detail.Based on the Mask R-CNN algorithm,the category information,positioning information and spatial layout pixel information of the target object in the input image can be extracted quickly and effectively.Meanwhile,image preprocessing is added to further improve the effect of image by using image segmentation,gray processing and other preprocessing methods.(2)For the style transfer task of the target object,based on the landscape style transfer method of convolutional neural network,we try to convert the target object area of the original image into an image with a specific style and we made it,moreover,we describe the key steps of the image in detail.Using VGG19 network to extract the feature in the first place,we extract content feature from high level,texture feature from low level,hence,characteristics is achieved.(3)For the image style fusion task,the research designs a two-stage image fusion scheme by analyzing the influences of different statistical factors on the fusion effect.Firstly,by information matching,preliminary integration happens.Secondly,with the help of preliminary design fusion,visual effect and fine adjustment model,the integration effect is improved.In the end,by redesigning the style of vector transmission standard and modifying calculation method of the loss function,the improve the quality of the reconstruction image style is achieved.