Research On Image Style Transfer Algorithm Based On Generative Adversarial Network

Painting is an ancient technique that dates back to the Paleolithic period,when ancient apes recorded their lives in the form of drawings.With the improvement of people’s material life,paintings are often used to preserve and convey the author’s message,and at the same time they bring pleasure to the senses.Drawing a particular style of image requires a certain level of skill,and deliberately learning to do so often takes a lot of time and experience.Image style migration,a branch of image processing,can transform a normal photo into an artistic style in a short time.The style migration technique can save the time cost of users,and is also widely used in the fields of beauty software and animation production.With the development of machine learning,the use of deep learning for image style migration has gradually achieved good results,but there are still some problems,such as not being able to take into account the local and overall effect,and the style performance is not obvious enough.In this thesis,based on generative adversarial networks,two different algorithms are proposed to improve the image style migration effect.The main research work of this thesis are summarized as follows:1.The existing style migration algorithms mainly decouple the content and style of the image,and then reorganize the content of the original image and the style of the reference image to generate an image with the target style,however,the boundary between the content and style of the image is rather blurred,and most of the algorithms define the content as the structural features of the underlying space and the style features as the rendering of the structure.Such a definition can clarify the target to be migrated,however,to a certain extent,it hinders the style migration,for example,the style at the content level is not migrated.In this thesis,we propose an unsupervised image style migration algorithm based on feature mapping,which enables feature mapping by improving the attention mechanism and realizes the mapping operation from the content of the original image to the content of the reference image,so as to realize the style migration of the image at the content level.The experiments show that the generated image can appropriately modify the content structure of the original image according to the content structure of the reference image,thus improving the effect of image style migration,while the relevant objective indexes also score higher.2.The algorithm of Research Work 1 can improve the effect of style migration,however,the content and style of the image cannot be decoupled well,so that the local information of the image can be effectively migrated,to address this problem,this thesis proposes an image style migration algorithm based on transfer learning,firstly,the model is pre-trained to separate the content and style of the image as much as possible,and then the formal model is trained,while at the same time,the content invariant loss is introduced,and the original content structure is still retained in the generated images,and finally,the effectiveness of the algorithm is demonstrated through comparison experiments.3.This thesis designs and implements an art painting style simulation system.This thesis encapsulates the algorithmic model proposed in Research Work 2 so that the theoretical research can be applied to practical problems,and the system is flexible enough to perform style migration and to be able to migrate the selected original image to the selected style image.