基于样式编码的真实图像逆映射算法  

A real image inverse mapping algorithm based on style coding

作　　者：尹芹[1] 方晖[1] 王金东[1] 王侃[1] 晏天文 霍智勇[2] YIN Qin;FANG Hui;WANG Jindong;WANG Kan;YAN Tianwen;HUO Zhiyong(Multimedia Video Products Department,ZTE Corporation,Nanjing 210000,China;School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]中兴通讯股份有限公司多媒体视讯产品部,江苏南京210000 [2]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2023年第1期80-85,共6页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：中兴通讯研究基金资助项目。

摘　　要：将图像准确地逆映射到StyleGAN的潜在空间,构造能够恢复真实图像的潜码,是实现真实图像语义操纵的基础。然而现有方法将图像逆映射为1×512维W潜码,重建后的图像并不能保真恢复。为此文中提出了一种基于样式编码网络的真实图像逆映射算法,其编码器网络提取粗、中、细不同的空间尺度特征作为风格特征向量,通过组合得到W+潜码后,送入预训练的StyleGAN生成器中,利用潜在空间重构出输入真实图像。实验表明,文中算法能够构造出真实图像在潜在空间的准确逆映射的潜码,提高了图像重构的质量,从而实现优于现有技术的图像语义操纵。Inversely mapping an image to the latent space of StyleGAN and constructing an accurate latent code that can recover the real image are crucial to the semantic manipulation of real images. However, the existing method only inversely maps the image into a 1×512 W latent code, and the reconstructed image cannot be restored with fidelity. In this regard, this paper proposes a real image inverse mapping algorithm based on a style coding network. The encoder network extracts three different spatial scale features, the coarse, the medium, and the fine, as style feature vectors. After the W+ latent code is obtained by combination, the input real images reconstructed from the latent space are fed into the pre-trained StyleGAN generator. Experiments show that the algorithm can better construct the latent code of the accurate inverse mapping of the real image in the latent space, improve the quality of image reconstruction, and realize image semantic manipulation than the existing technology.

关 键 词：潜在空间 逆映射 样式编码 语义操纵 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]