Makeup Based On Image Segmentation And Local Transfer

With the wide application of face interaction system,the security and reliability of face interaction system have become a hot issue.Face forgery prevention is one of the important areas,the goal is to recognize faces is real or fake images,forge the way including the true face of bogus(impersonation attack)and confused face fake(obfuscation attack),and this the real face of fake fake pictures,fake video,confused face of bogus includes facial make-up and shade.Although promising progress has been made,the existing work still has difficulties in dealing with complex spoofing attacks and in generalizing to real world scenarios.The main reason is that the existing anti-forgery data sets are limited in quantity and diversity.In order to overcome these obstacles,in addition to the relatively easy collection of fake data of non-real faces,this paper first introduces an automatic makeup imitation system based on semantic segmentation and local migration.The user provides source and target images for reference.The system takes the target image as a reference and simulates the relevant makeup of the target image in the designated area of the source image.More accurately,we put the target image each for makeup,lip gloss,eye shadow,eyebrows)transfer to the corresponding area of the source image face,thereby gaining new images of the fake target human face,the experimental results show the effectiveness of the method,based on the above method,we can get a lot of confusion face this kind of new type face counterfeit means data.Based on the above algorithmic synthesis and manual acquisition,we constructed a large scale face anti-spooking data set,Celeba-Spoof,with the following attractive properties: Fake images were captured from 8 scenes(2 environments * 4lighting conditions)and over 10 sensors.Celeba-Spoof contains annotations for 10 face forgery types as well as 40 face attribute annotations inherited from the original Celeba-Spoof dataset.We are equipped with Celeba-Spoof to carefully benchmark existing face anti-counterfeiting approaches in a unified multitasking framework –the Auxiliary Information Embedding Network(AENet)and reveal some valuable observations.Our key finding is that rich semantic tagging,as an auxiliary task,can greatly improve performance and generalization against forgery attacks compared to the commonly used binary supervised or middle-level geometric representations.Through comprehensive research,we found that Celeba-Spoof is an effective training data source.The model trained on Celeba-Spoof(without fine-tuning)shows state-of-the-art performance on standard benchmarks such as CASIA-MFSD.