Research On Scene Text Editing Technology Based On Generative Adversarial Networks

Scene text editing refers to replacing the text in an image with the target text of the same style while preserving the background as much as possible.Because of its wild applications in text and image synthesis,advertising design,image restoration,text information hiding,AR translation,etc.,this technology attracted a lot of research interests in the past few years,and some significant progress has been achieved.For example,the generative adversarial network based SRNet proposed by Wu et al.performs fairly well for word-level scene text editing task.However,in many practical applications,scene text editing is still challenging,because the background texture is complex usually,the text style is hardly to capture accurately,and the target text may be not as long as the original one.This thesis proposes the SRNet+MS network.Same as SRNet,the network is generative adversarial network based,and consists of three sub-networks:the background inpainting network,the text style transfer network,and the image fusion network.The differences from the SRNet include:(1)A mask attention(MA)module based on non-local connection operations is introduced in the background restoration network,which aims to build long-range dependences among pixels,so as to achieve more accurate background inpainting through mask constraints.(2)In the text style transfer network,some spatial transformation(ST)modules are adopted as a substitution of the skeleton-guided learning mechanism that SRNet used,so as to well capture complex spatial structure of scene text.After being trained on a synthetic data set,the SRNet+MS was tested on the real data set ICDAR13/17 without any adjustments.Intensive experiments show that the network performs well for the tasks of scene text editing,image inpainting and visual translation,which is shown to be superior to the SRNet.