结合多注意力及IE-GAN的面部属性编辑方法
A facial attribute editing method combining multi-attention and IE-GAN
石晓楠 1息佳琦 1王英丽1
作者信息
- 1. 黑龙江大学 电子工程学院,哈尔滨 150080
- 折叠
摘要
面部属性编辑主要有两个目的:①将图像从源域转换到目标域,对指定的目标属性(如性别、年龄、发色等)进行修改;②只改变与目标属性相关的人脸区域,保留目标属性之外其他属性的细节.而现有的人脸属性编辑方法不可避免地会对与目标属性无关区域进行更改.因此,提出一种基于 IE-GAN 和多注意力机制的面部属性编辑方法 MAIE-GAN,引入补充注意连接(Complementary attention connection,CAC)的概念连接编解码器,解决了由直接跳跃连接引起的信息冗余问题,将自注意力机制作为生成器中卷积层的补充,使其能更好定位目标属性,限定属性变换区域,此外,还利用互补注意特征的概念,实现目标属性无关区域的更好保留.通过与现有方法的对比分析表明,该方法在属性定位和图像质量方面优于现有的方法.
Abstract
Face attribute editing has two main purposes:① to transform the image from the source domain to the target domain and make changes to the specified target attributes(such as gender,age,hair color,etc.);② to change only the face regions related to the target attributes and retain the details of other attributes outside the target attributes.And the existing face attribute editing methods inevitably make changes to the regions unrelated to the target attributes.Therefore,a face attribute editing method MAIE-GAN based on IE-GAN and multi-attention mechanism is proposed,introducing the concept of complementary attention connection(CAC)to connect codecs,solving the information redundancy problem caused by direct jump connection,the self-attention mechanism is used as a complement to the convelutional layer in the generator to enable better localisation of the target attributes and confinement of the attribute transformation region,in addition to using the concept of complementary attention features to achieve better retention of irrelevant regions of the target attributes.A comparative analysis with existing methods shows that this method outperforms existing methods in terms of attribute localization and image quality.
关键词
面部属性编辑/空间注意力机制/生成对抗网络/补充属性连接Key words
facial attribute editing/spatial attention mechanism/generative adversarial network(GAN)/complementary attention connection引用本文复制引用
基金项目
黑龙江省自然科学基金(YQ2020F012)
装备重大基础研究项目(514010??????)
出版年
2024
NETL
NSTL
万方数据