Research On Opera Makeup Transfer With Generative Adversarial Network

Makeup transfer is a hot research direction in the forefront of computer vision,which belongs to the category of style transfer and face processing,and aims to transfer any target makeup to a specified face image.Makeup transfer not only includes face alignment and color transfer,but also takes into account the light,shadow and texture differences of different faces.A large number of scholars have carried out research on makeup transfer tasks,and have achieved certain results,but there are still some challenges.For example,it is difficult for the transfer model to deal with makeup transfer with a large style gap at the same time;the algorithm has high requirements on image quality,and differences in shadows,occlusions,and posture expressions will reduce the quality of makeup transfer;image feature entanglement has a greater impact on the effect of makeup transfer,and so on.In view of the above problems,the main research contents of this thesis include:1.Aiming at the difficulty of traditional makeup transfer algorithms to deal with makeup transfer problems with large style differences,a makeup transfer algorithm(Adaptive Cycle Generative Adversarial Network)based on adaptive instance normalization is proposed.ACGAN takes the recurrent generative adversarial network as the framework and introduces adaptive instance normalization technology.The model can actively learn the styles of different makeup under the premise of ensuring the identity of the makeup.The experimental results on the cross-style makeup datasets show that the improved network has improved learning ability for different styles of makeup,and can better handle makeup transfer tasks.2.Aiming at the influence of different pose expressions and shadows on the makeup transfer effect in reality,a makeup transfer algorithm(Self-attention ACGAN,SACGAN)based on selfattention is proposed.Based on ACGAN,the algorithm model adds a self-attention restoration module and a attention-makeup morphing module to perform shadow repair and deformation alignment on face makeup to enhance the makeup transfer effect of contaminated images.The experimental results show that SACGAN reduces the influence of pose inconsistency and image shadow on the transfer effect,and improves the robustness of the algorithm.3.Aiming at the problems of color distortion and makeup smearing in the makeup transfer caused by feature entanglement,combined with the characteristics of opera makeup,a makeup transfer algorithm(Style-disentangling ACGAN,Style-ACGAN)based on latent space disentangling is proposed.The latent space disentangling of image features is carried out through the mapping network and adaptive instance normalization to guide the completion of the makeup transfer task and reduce the impact of feature entanglement on the transfer effect.The experimental results show that the introduction of feature disentangling enhances the color transfer ability of the algorithm and further improves the effect of makeup transfer.