选择性传输与铰链对抗的多图像域人脸属性迁移

在基于生成对抗网络的人脸属性迁移过程中，存在图像域表达形式单一、图像域迁移细节失真的问题。提出一种结合选择性传输单元与铰链对抗损失的多图像域人脸属性迁移方法。在生成器中，利用自适应实例归一化融合图像的内容信息与图像域控制器生成的样式信息，增加图像域表达方式的多样性，同时通过选择性传输单元将下采样提取的内容特征根据相对属性标签选择性地传输到上采样，形成融合特征以增强图像的细节信息。在判别器中，通过增加双尺度判别，协同鉴定人脸图像的真伪及类别，从而提高判定的准确度。在此基础上，设计融合相对鉴别和铰链损失的对抗损失函数，增强真伪图像域之间的联系。在CelebA数据集上的实验结果表明，与StarGAN、STGAN等主流的多图像域人脸属性迁移方法相比，该方法能够建立更准确的多图像域映射关系，提高迁移图像的质量同时增加迁移图像表达的多样性。

Facial Attribute Migration Across Multiple Image Domains Using Selective Transfer and Hinge Adversary

In facial attribute migration based on a Generative Adversarial Network(GAN), the expression form of the image domain is single and the details of the image domain migration are distorted.This paper presents facial attribute migration across multiple image domains method, which combines a selective transfer units and a hinge against loss.In the generator, the adaptive instance normalization and fuse the content information of the image and the style information that is generated by the image domain controller to increase the diversity of the expression methods in the image domain.The content features that are extracted from the downsampling are selectively transmitted to the upsampling according to the relative attribute label through the selective transmission unit to form fusion features for enhancing the detail information of the image.In the discriminator, dual-scale discrimination is included to identify the authenticity and category of face images jointly, which improves the judgment accuracy.On this basis, an anti-loss function that integrates relative discrimination and hinge loss is designed to enhance the relationship between the true and false image domains.The experimental results on the CelebA dataset demonstrate that, compared to mainstream facial attribute migration across multiple image domains methods such as StarGAN and STGAN, this method can establish a more accurate multiple image domains mapping relationship, improve the quality of the migrated image, and increase the diversity of the expression of the migrated image.