基于生成对抗网络的自动美妆技术研究

Research on Automatic Makeup Technology Based on Generative Adversarial Network

现有的基于生成对抗网络的自动美妆方法存在着生成的图像质量不高、美妆效果不理想等问题。因此,论文研究了一种WarpMakeup-GAN方法实现了女性人脸图像的自动美妆功能。首先,对于网络结构部分,通过深度可分离卷积和特征金字塔结构提升网络速度和拟合能力;其次,对于损失函数部分,通过自重建损失函数和仿射变换损失函数提升网络生成的图像质量和美妆效果;最后,通过测试阶段引入人脸对齐进一步优化美妆效果。与现有的方法相比,此方法可以生成更好的美妆图像。

Existing automatic makeup methods based on generative adversarial network have problems such as low quality of generated images and unsatisfactory makeup effects.Therefore,this paper studies a WarpMakeup-GAN method to realize the automatic makeup function of female face images.First,for the network structure part,the network speed and fitting ability are improved through the depthwise separable convolution and feature pyramid structure.Secondly,for the loss function part,the image quality and makeup effect generates by the network were improved through the self-reconstruction loss function and the affine transformation loss function.Finally,the face alignment is introduced through the test phase to further optimize the makeup effect.Compared with existing methods,this method can generate better beauty images.