基于生成对抗网络的人脸妆容迁移方法研究
Research on Face Makeup Transfer Method Based on Generative Adversarial Networks
孙克雷 1潘宇 1童波1
作者信息
- 1. 安徽理工大学计算机科学与工程学院,安徽淮南,232000
- 折叠
摘要
妆容迁移是一项利用计算机视觉和深度学习算法将一种妆容的风格转移到其他人脸上的技术,以实现高仿妆效果转换.为了有效解决现有人脸妆容迁移方法中存在的上妆区域错误和妆容迁移不完整的问题,提出了一种基于生成对抗网络的人脸妆容迁移方法(MutNet).以解决妆容迁移不完整的问题为目标,该方法在解码器中引入了空间注意力机制,来帮助网络更加聚焦于需要修改的区域,并通过引入孪生对比损失,更好地实现人脸之间的语义对应关系,有效缓解或克服上妆区域错误的问题.同时通过与其他方法的对比结果表明,MutNet能获得更协调的上妆效果.
Abstract
Makeup transfer is a technique in computer vision and deep learning that involves transferring the style of one makeup look to other faces to achieve a high-fidelity makeup effect transformation.To effectively address the issues of makeup transfer,such as makeup region errors and incomplete makeup transfer,a makeup transfer method based on Generative Adversarial Networks(GANs),known as MutNet,is proposed.With the goal of addressing in-complete makeup transfer,this method introduces a spatial attention mechanism in the decoder to help the network focus more on the areas that need modification;incorporating Siamese contrastive loss to better establish semantic correspondences between faces,it effectively mitigates or overcomes makeup region errors.At the same time,the comparative results with other methods show that MutNet can achieve a more coordinated makeup effect.
关键词
人脸妆容迁移/人脸图像生成/生成对抗网络/孪生对比损失/空间注意力机制Key words
Face makeup transfer/Face image generation/Generative adversarial networks/Siamese contrastive loss/Spatial attention mechanism引用本文复制引用
基金项目
国家自然科学基金项目(62076006)
安徽省高校重点科研项目(2022AH050821)
安徽理工大学研究生创新基金项目(2023cx2126)
出版年
2024
国家科技期刊平台
NSTL
万方数据