Face Editing Based on Facial Recognition Features
Face Editing Based on Facial Recognition Features
复制标题
DOI：
10.1109/tcds.2022.3182650
复制
发表时间：
2023-06-01
影响因子：
5
通讯作者：
Jiang, Yizhang
中科院分区：
文献类型：
--
作者：
Ning, Xin;Xu, Shaohui;Jiang, Yizhang
人脸编辑生成具有目标属性的人脸图像，而不改变身份或其他信息。现有的人脸识别方法虽然取得了较好的识别效果，但在控制属性强度的同时，不能有效地保留人脸的身份和语义信息。受两个人类认知特征的启发，即全局优先原则和同源连续原则，我们提出了一种新的人脸编辑方法，称为信息保留和强度控制生成对抗网络（IricGAN）。它包括一个可学习的分层特征组合（HFC）功能，可以通过多尺度特征混合来构建样本的源空间;它可以在显著压缩网络的同时保证源空间的完整性。此外，属性回归模块（ARM）可以解耦源空间中的不同属性范式，以确保正确修改所需的属性，并保留其他区域。可以通过在源空间中应用不同的控制强度来模拟修改面部属性的渐进过程。在人脸编辑实验中，定性和定量结果都表明，IricGAN在最先进的替代方案中取得了最好的整体效果。通过对源空间和图像的关系进行反馈，目标属性可以不断修改，并且最大程度地保持了各属性的独立性。IricGAN：https://github.com/nanfangzhe/IricGAN。
Face editing generates a face image with the target attributes without changing the identity or other information. Current methods have achieved considerable performance; however, they cannot effectively retain the face's identity and semantic information while controlling the attribute intensity. Inspired by two human cognitive characteristics, namely, the principle of global precedence and the principle of homology continuity, we propose a novel face editing approach called the information retention and intensity control generative adversarial network (IricGAN). It includes a learnable hierarchical feature combination (HFC) function, which can construct a sample's source space through multiscale feature mixing; it can guarantee the integrity of the source space while significantly compressing the network. Additionally, the attribute regression module (ARM) can decouple different attribute paradigms in the source space to ensure the correct modification of the required attributes and preserve the other areas. The gradual process of modifying the face attributes can be simulated by applying different control strengths in the source space. In face editing experiments, both qualitative and quantitative results demonstrate that IricGAN achieves the best overall results among state-of-the-art alternatives. Target attributes can be continuously modified by refeeding the relationship of the source space and the image, and the independence of each attribute can be retained to the greatest extent. IricGAN:https://github.com/nanfangzhe/IricGAN.