Research On GAN-based Image De-discrimination Synthetic Method

With the rapid development of the Internet and the continuous updating of social media,image sharing has become the main function of various social network platforms.By sharing images,people can get information and knowledge.But many images on social media spread discriminatory attitudes consciously or unconsciously.The emergence and dissemination of such images affect the relationship between people.What's worse,it has a negative impact on social harmony and the growth of the youth.At the same time,image discrimination has diversity.It is of great significance to detect sensitive areas in images and remove discriminatory areas on the basis of guaranteeing the basic semantics of images,that is,to study non-discriminatory image synthesis methods.Aiming at the body language discrimination in discrimination,the main work of this thesis is as follows:(1)The discrimination in body language is divided into eight categories,and the non-discrimination content that interferes with discrimination is divided into eight categories,according to the current mainstream categories of discrimination in body language,and the analysis of two forms of discrimination:face-to-face and back-to-face,The data set of discrimination and non-discrimination is constructed independently,and the data labeling is completed,including frame selection of the human body,face and hand in the data set.(2)There are many types of discrimination in data concentration,and the discrimination relationship is often composed of multiple objects such as gestures and expressions.Only detecting a single object can lead to the occurrence of misdetection results.Therefore,traditional target detection methods cannot be sensitive to discrimination.In this thesis,a method of target detection for discriminatory sensitive objects is proposed,which combines Faster R-CNN with ResNet network and can effectively detect discriminatory sensitive objects.The validity of the model is verified by Class Activation Map(CAM).The redundant images in the preliminary test results were removed,and the effectiveness of the method was verified by experiments.(3)In this thesis,we propose expression replacement method based on CycleGAN and gesture replacement method based on InstaGAN,which can effectively replace discriminatory sensitive objects.Through experiments,the effectiveness of alternative methods for discriminating sensitive objects was verified.