Text Remove Network Based On GAN And Its Application

As an important carrier of daily information transmission,text pictures are ubiquitous in our daily lives.These pictures usually contain a large amount of personal privacy information,such as phone numbers,home addresses,ID numbers and various account information.Once criminals obtain pictures containing this information,it is easy to leak privacy and cause serious personal property loss.Recently,crimes caused by privacy leaks have occurred frequently.Therefore,in order to prevent the leakage of text-sensitive information,a variety of text removal methods have been proposed in recent years.These methods can remove the sensitive text in the picture to a certain extent,but there are still some problems such as unclean text erasure and serious background damage.This article mainly focuses on these issues to study the text removal algorithm.The main work includes the following aspects:1.This paper proposed a GAN-based text removal network-TRNet.By optimizing the generator,discriminator and comprehensive loss function configuration,it solved the problems of incomplete text removal,color distortion,loss of background texture,and embossment in the text area;2.We designed a selectable area text removal method based on TRNet,which can remove sensitive text while retaining insensitive text.At the same time,we designed a discriminator model of patch 100 to solve the problem of unclean removal of large fonts;3.In order to simplify the data labeling process and enhance training dataset distribution,we proposed an end-to-end text synthesis method,which can synthesize text-sensitive area text datasets end-to-end,and the synthesized data is used as a supplement to the text removal training dataset,which greatly improves the performance of the text removal method.At the same time,we also optimized the synthesis of the training set to solve the artifact problem caused by the large pixel gradient of the text edge to a certain extent.