基于特征向量变换GAN的多域图像转换方法  

Multi-domain image conversion method based on feature vector transformation GAN

作　　者：王思明[1] 李昭坊 WANG Si-ming;LI Zhao-fang(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《云南大学学报（自然科学版）》2020年第6期1080-1090,共11页Journal of Yunnan University(Natural Sciences Edition)

基　　金：国家自然科学基金(61867003,61263004).

摘　　要：生成对抗网络(Generative Adversarial Nets,GAN)在图像翻译及多域图像转换方向已取得显著成功,但现有用于多域间图像转换的GAN大部分使用了多个生成器G及判别器D,导致网络训练参数量过大,数据集不能充分利用.针对以上问题,基于StarGAN和多模态无监督图像转换方法,提出基于特征向量变换的GAN模型.首先,将源图像编码成内容向量加特征向量的形式;然后将提取到的特征向量从源域转换到目标域而保持内容向量不变;最后完成图像重构.该模型仅使用一对生成器G和判别器D,有效地解决了上述问题.相较于现有模型,新模型不仅适用于多域图像转换,还可以从噪声生成指定图像.在CelebA数据集上的实验结果表明新模型与现有模型相比在多域人脸属性转换方面表现出更好的效果.The Generative Adversarial Nets(GAN) has achieved remarkable success in image translation and multi-domain image conversion.However,most of the existing GANs used for multi-domain image conversion use multiple generators G and discriminators D,resulting in an excessive amount of network training parameters and insufficient utilization of data sets.In view of the above problems,we propose a GAN model based on feature vector transformation,which based on StarGAN and multi-modal unsupervised image conversion method.Firstly,the model encodes the source image into a form of a content vector and a feature vector.Then,the model converts the extracted feature vector from the source domain to the target domain while the content vector remains unchanged.Finally,the image reconstruction is completed.This model solves the above problems effectively by using only one pair of generator G and discriminator D.Compared with the existing models,this model is not only suitable for multi-domain image conversion,but also for generating the specified image from noise.The experiments on the CelebA dataset show that the proposed model performs better in terms of multi-domain face attribute conversion than existing models.

关 键 词：生成对抗网络 特征向量 图像转换 图像生成 CelebA数据集 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]