Image style transfer is the task of transferring the style style to the target area in the source image to create artistic typesetting.This paper studies how to transform the style of the text area in the scene text image,in order to achieve automatic adver-tising or posters in the text for style conversion,reduce the cost of artistic creation and improve the diversity of artistic style.Due to the complex interaction between different factors in the scene text image,there are few previous work on text style conversion while retaining the original text content and background.This paper presents a three-stage framework which is the first network to carry out a degree-controlled style conversion directly in the original image,the method of style conversion for single binary character is extended to the text on the scene text image,and the related knowledge of image restoration is involved.First,the style conversion network is used to transform only the text style in the scene text image,and then the original character is erased by the character erasing network to reconstruct the background image.Finally,the fusion part uses the generated foreground image and the back-ground image after erasing the characters to generate the final image after the style conversion.In this paper,a large number of ex-periments are carried out to prove the effectiveness of the method.
deep learningGANscene text imageimage style transferfont style tranfercharacter erasure
国家科技期刊平台
NSTL
万方数据
