Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (4): 208-215.DOI: 10.3778/j.issn.1002-8331.2110-0223

• Graphics and Image Processing • Previous Articles     Next Articles

Controllable Face Image Synthesis Algorithm Based on Attribute Decomposition and Fusion

LIANG Hong, CHEN Qiushi, SHAO Mingwen   

  1. College of Computer Science and Technology, China University of Petroleum(Hua Dong), Qingdao, Shandong 266000, China
  • Online:2023-02-15 Published:2023-02-15

基于属性分解融合的可控人脸图像合成算法

梁鸿,陈秋实,邵明文   

  1. 中国石油大学(华东) 计算机科学与技术学院,山东 青岛 266000

Abstract: In real life, it is difficult to directly collect face images due to privacy or security factors, so image generation methods can be considered. When using a generative adversarial network for image generation, the results are prone to problems such as low resolution, blurred edges, and loss of identity information features. In response to the above problems, this paper proposes a new face feature generation model:by embedding key information as an independent code into the implicit space, and fusing and interpolating with global features, achieve the controllable generation of key facial features; introduce the improved attention module, and pay attention to the correlation between local features and global features during the generation process; introduce the color difference loss and face component loss into the overall loss function, and responsible for constraining pixel color and facial texture features. The algorithm can generate natural and true appearance features in local areas of the face, retain the original identity information, and generate smooth facial contours. Experiments using the preprocessed CelebA dataset show that the algorithm has a significant improvement in subjective visual effects, and at the same time has a stable improvement in PSNR and SSIM compared with existing methods.

Key words: generative adversarial networks, face feature generation, attention mechanism

摘要: 在现实生活中,人脸图像受隐私或安全因素的限制难以直接采集,因此可以考虑采用图像生成方法。当使用生成对抗网络进行图像生成时,容易出现分辨率低、边缘模糊、身份信息特征丢失等问题。针对上述问题,提出了一种新的人脸特征生成模型:通过将关键信息作为独立编码嵌入隐式空间,再与全局特征进行融合插值实现对人脸关键特征的可控生成;引入改进的注意力模块,在生成过程中关注局部特征和全局特征的相关性;将色差损失和人脸分量损失联合引入整体损失函数中,负责约束像素颜色和人脸纹理特征。该算法可以在人脸局部区域生成自然真实的外观特征,保留原始身份信息,并生成平滑的面部轮廓。使用预处理后的CelebA数据集的实验表明,该算法在主观视觉效果上有显著提升,同时与现有方法相比在PSNR和SSIM上有稳定的提升。

关键词: 生成对抗网络, 人脸特征生成, 注意力机制