摘要
针对较为复杂的人物图像生成任务,提出了一种从mask图像生成人物图像的新方法。基于生成式对抗网络(generative adversarial nets,GANs),接收一个多通道的mask图像作为输入,其中的每个通道表示人物某个区域(如头发、脸部、手臂等)的掩码。该网络由生成器和判别器组成,生成器在U-Net结构基础上加入了残差模块,判别器用于判别生成图像的真伪性。通过施加不同的高斯噪声,所提方法能根据相同的mask生成具有不同外观的人物图像,具有更好的结果多样性。
To fulfil the complicated task of human image synthesis,in this paper we propose a method to synthesize human images from masks.The proposed method is designed based on conditional generative adversarial nets(GANs),and the networks receive a multi-channel mask image as input,each channel of which is a binary mask indicating existence of one part of human body(hair,face,arms,etc.).The network contains a generator and a discriminator.The generator is a U-Net structure with residualblocks,and the discriminator is incorporated to identify authenticity of the generated human image conditioned on its mask.Human images with varying appearance could be generated from the same input mask by adding different Gaussian noise.
作者
欧阳雯琪
徐昆
OUYANG Wenqi;XU Kun(Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China)
出处
《中国科技论文》
CAS
北大核心
2019年第3期255-260,共6页
China Sciencepaper
基金
国家自然科学基金资助项目(61822204)
关键词
生成式对抗网络
生成器
判别器
人物图像生成
generative adversarial nets
generator
discriminator
human image synthesis
