基于STGAN的人脸属性编辑改进模型  

Improved model of face attribute editing based on STGAN

作　　者：林志坤 许建龙[2] 包晓安[2] LIN Zhikun;XU Jianlong;BAO Xiao'an(School of Information Science and Engineering,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of Computer Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息科学与工程学院,杭州310018 [2]浙江理工大学计算机科学与技术学院,杭州310018

出　　处：《浙江理工大学学报（自然科学版）》2023年第3期285-292,共8页Journal of Zhejiang Sci-Tech University(Natural Sciences)

基　　金：浙江省重点研发计划项目(2020C03094)。

摘　　要：人脸属性编辑在美颜APP和娱乐领域有重要应用,但现有方法存在生成图像质量不高、属性编辑不够准确等问题,为此提出了一种基于选择传输生成对抗网络(Selective transfer generative adversarial networks, STGAN)的人脸属性编辑改进模型。运用潜码解耦合思想,将潜码分解为内容潜码和风格潜码单独操作,提高源域图像和目标域图像的内容编码一致性,从而提高属性编辑准确率;同时运用像素级重构损失和潜码重构损失,在总损失函数中加入像素级限制和潜码重构限制,通过互补作用提高生成图像质量。在CelebA人脸数据集和季节数据集上进行实验,该模型相比当前人脸属性编辑主流模型在定性结果和定量指标上均有提高,其中峰值信噪比和结构相似性相比STGAN模型分别提高了6.06%和1.58%。这说明该改进模型能够有效提高人脸属性编辑的性能,满足美颜APP和娱乐领域的需求。Face attribute editing technology has important applications in beauty APPs and entertainment fields.However,the existing methods still have problems such as low-quality and inaccurate editing.To this end,an improved face editing model based on selective transfer generative adversarial networks(STGAN)was proposed.Using the idea of latent code decoupling,the latent code was decomposed into the content latent code and the style latent code,which improved the content-coding consistency of the source domain image and the target domain image,thereby improving the accuracy of attribute editing.In the meanwhile,we used pixel-level reconstruction loss and latent code reconstruction loss,and added pixel-level restrictions and latent code reconstruction restrictions to the total loss function,improving the quality of generated images through complementary effects.Experiments were carried out on the CelebA face dataset and seasonal dataset.Compared with the current mainstream model of face attribute editing,this model has improved both qualitative results and quantitative indicators.Compared with the STGAN model, the peak signal-to-noise ratio and structural similarity index of this model areimproved by 6.06% and 1.58%, respectively. This shows that the improved model can effectivelyimprove the performance of face attribute editing and meet the needs of beauty apps and entertainment fields.

关 键 词：生成对抗网络 人脸编辑 重构图像 潜码解耦 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]