Research And Implementation Of Facial Makeup Transfer Based On Generative Adversarial Networks

In recent years,the synthesis and analysis of face images has become a hot research topic in the field of computer vision.Face makeup transfer is to transfer the makeup of the reference image to the source image(uncolored image),which is a kind of face image synthesis and analysis,and it has huge application prospects in film and television entertainment,beauty industries,etc.With the development of Generative Adversarial Networks(GAN),face makeup transfer model based on generative adversarial network has achieved milestone progress.However,due to the need to not only face cosmetic components extracted from the reference image,but also to analyze the facial structure and consider the image posture,expression,illumination and other factors in the process of face makeup transfer,the current face makeup transfer algorithm based on generative adoration network still faces many problems and challenges in application and theory.Based on extensive reading of existing references,this paper analyzes the application scenarios,advantages and disadvantages of the existing makeup transfer methods,and studies the face makeup transfer based on generative adversarial network for the accuracy and robustness of face makeup transfer.The main research contents and innovations are as follows:(1)Face makeup transfer algorithm based on self-attention mechanism.In order to achieve a more accurate face makeup transfer effect,a face makeup transfer algorithm based on the self-attention mechanism is proposed.The self-attention mechanism module is introduced into the generative adversarial network,and its generator and discriminator are improved.Firstly,in order to reduce the loss of image information during the transfer process,the generator in this paper adopts a U-net structure,and the encoder and decoder use a residual structure.Adding a self-attention mechanism to the second-layer residual block of the decoder can strengthen the ability of the model to capture features.Then the discriminator adopts the improved Markov discriminant model to enhance the ability to discriminate the details of the generated image.Finally,an identity preservation loss and a background invariant loss are introduced to help synthesize realistic and accurate facial makeup transfer images.The experimental results show that this algorithm can improve the quality of makeup transfer and generate more accurate transfer results while maintaining identity information and background information.(2)Face makeup transfer algorithm based on 3D perception spatial transformation.In order to improve the robustness of face makeup transfer algorithm,a new face makeup transfer algorithm based on 3D perception spatial transformation is proposed.Given the source image and the reference image,a 3D face model is first fitted to decompose the face into shapes and textures.In the texture branch,map the texture to UV space and design a space transform generator to transfer the makeup effect.The spatial transformation generator includes a spatial transformation network module and a makeup transfer module.The spatial transformation network module aligns the source texture features with the reference texture features,and the makeup transfer module makes the makeup transfer results more accurate.This algorithm can eliminate the difference in pose and expression of the input images.The experimental results show that the proposed algorithm can further improve the robustness of face makeup transfer.There are 27 pictures,5 tables,and 87 reference in this thesis.