Design And Implementation Of Arbitrary Face Exchange Algorithm Based On Generative Adversarial Networks

Face swapping is to replace the identity of a person in the target image with the identity of another person in the source image,while preserving the original attributes,such as head posture,facial expression,skin color,lighting,background,etc.Face swapping has wide application value in privacy protection,video synthesis,game rendering,and other creative applications.Traditional face swapping often relies on a lot of manual labor to eliminate defects,there are a lot of repeated operations,low efficiency,and poor general performance.A universal,stable and efficient face swapping algorithm is facing increasing demands.In recent years,with the development of graphics and deep learning,the effect and efficiency of image generation have been greatly improved.Among them,generative adversarial networks have achieved great success in the field of image generation,and the research on generative adversarial networks has also achieved very rich results.Aiming at the problem of poor replacement effect of several current mainstream face swapping algorithms,this paper proposes a single model and single training based on generating an adversarial network,cross-regrouping three-way features of identity,attributes,and facial feature points.A single reference face image can realize the algorithm of any face replacement in the open set,and around this algorithm is designed and built an automatic framework that can stably and efficiently perform batch face swapping.The main work is as follows:(1)The structure and process of any face swapping algorithm are proposed,and the improved WGAN is used as the basic structure of the face swapping algorithm;an identity encoder,attribute encoder,and facial feature point extractor are introduced to couple identities to facial feature points.The problem that features affect the generator's identity retention ability,a facial feature point conversion module is proposed to decouple the identity information in the facial feature points,and maximize the role of feature points in maintaining posture and expression;the loss function and training process of the face swapping algorithm is introduced,thus the model can self-supervise the training of the attribute encoder and generate high-quality replacement results.(2)Besize the face swapping algorithm,a set of stable and efficient automated frameworks for batch face replacement is proposed.The pre-processing process extracts the face in the scene image and aligns it with the standard face,reducing the influence of the background and angle;post-processing uses inverse transform to transform the replacement face back to the original scene image position,and uses Poisson fusion to replace Faces fit seamlessly back to the original scene image.Aiming at the halo problem caused by Poisson fusion,a strategy for generating segmentation templates is proposed,which can eliminate the halo caused by Poisson fusion to the greatest extent.(3)Put forward quantitative indicators to measure the quality of the replacement face image,identity retention ability and attribute retention ability,and on this basis,a lot of comparative experiments were done.Through ablation experiments on each module of the face replacement algorithm,it is proved that the facial feature point extraction module can significantly improve the ability to maintain the posture and expression of the generated image.At the same time,the feature point conversion module proposed in this paper can effectively decouple the facial feature points.Identity information and posture expression information,while ensuring that the generated image has high posture expression consistency,it also has very high identity consistency;this paper also on the fake face video detection benchmark data set Face Forensics ++,and the existing mainstream face replacement algorithm The qualitative and quantitative comparison of Deepfakes,faceswap and MUNIT shows that the algorithm in this paper is superior to the current mainstream face replacement method in generating image quality,identity retention ability,posture and expression retention ability,proving the correctness and superiority of the algorithm.