Research On Text Editing Technology Based On GAN

Text play an extremely important role in the development of human.As a bridge of civilization and communication,the emergence of text has a tremendous impact on work and life of human.In recent years,with the development of computer vision technology,the text has also attracted more and more attention from researchers.The research of scene text detection and recognition,text style transfer and text erasure are also developing constantly,but in the current research,there is little work on the direction of the text editing.Text editing has a very large application in the work and life,like information hiding,poster advertising reuse or visual scene translation such as AR translation,etc.,all need to replace the text in the picture,but the general method to accomplish this task is using tools such as PS,and it is not simple and efficient.This paper takes scene text editing which is a new research direction as the main goal of the research.The text editing task refers to given a scene text image and text to be replaced,a new image which retains the style of original scene text image and the content of input text can be generated through the network,and the generated image is guaranteed to be realistic.This paper proposes a new network called SRNet for text editing task,which can perform well on text editing and word-level text erasure at the same time.Our network is mainly composed of three parts,namely,foreground text transfer network,background erasure network,and fusion network.The foreground text transfer network is responsible for assigning text-related style feature,such as font,color and deformation to the new text,the background erasure network erases the text in the original style text image and complementing the visual sensory texture information,the final fusion network merge the transferred foreground text image and the processed background to ensure its visual realistic.The main work and contribution of this paper as follows:1)The network proposed in this paper that can complete the word-level editing task,and also can achieve end-to-end training and prediction;2)The innovative proposal of this paper is sperating the text editing task into modular sub-tasks to achieve a more realistic effect;3)The method of this paper has achieved excellent performance in tasks related to text editing,such as text editing in the same language,cross-language text editing,and text erasure.