基于多尺度 Scale-Unet 的单样本图像翻译-《计算机技术与发展》

文章信息/Info

作者:: 周蓬勃¹ ; 冯龙^2* ; 寇宇帆²; 1. 北京师范大学艺术与传媒学院,北京 100032;
2. 西北大学信息科学与技术学院,陕西西安 710127

Author(s):: ZHOU Peng-bo1 ; FENG Long2* ; KOU Yu-fan2; 1. School of Art and Media,Beijing Normal University,Beijing 100032,China;
2. School of Information Science and Technology,Northwest University,Xi’an 710127,China

Keywords:: single-sample image translation; Scale-Unet; multi-scale structure; progressive approach; scale-pixel loss

摘要:: 随着生成对抗网络(GAN) 的发展,基于单样本的无监督图像到图像翻译(UI2I) 取得了重大进展。然而,以前方法无法捕获图像中的复杂纹理并保留原始内容信息。为解决这个问题,提出了一种基于尺度可变 U-Net 结构( Scale—Unet)的新型单样本图像翻译结构 SUGAN。所提出的 SUGAN 使用 Scale—Unet 作为生成器,利用多尺度结构和渐进方法不断改进网络结构,以从粗到细地学习图像特征。同时,提出了尺度像素损失 scale-pixel 来更好地约束保留原始内容信息,防止信息丢失。实验表明,与 SinGAN、TuiGAN、TSIT、StyTR2 等公共数据集 Summer－Winter、Horse－Zebra 上的方法相比,该方法生成图像的 SIFID 值平均降低了 30% 。所提方法可更好地保留图像内容信息,同时生成详细逼真的高质量图像。

Abstract:: Single-sample unsupervised image - to - image translation ( UI2I ) has made significant progress with the development ofgenerative adversarial networks ( GANs) . However,previous methods cannot capture complex textures in images and preserve originalcontent information. We propose a novel one - shot image translation?
structure SUGAN based on a scale - variable U - Net structure( Scale—Unet) . The proposed SUGAN uses Scale—Unet as a generator to continuously improve the?
network structure using multi-scalestructures and progressive methods to learn image features from coarse to fine. Meanwhile, we propose the scale - pixel loss to?
betterconstrain the preservation of original content information and prevent information loss. Experiments show that compared with SinGAN,TuiGAN,TSIT,StyTR2 and?
another methods on public datasets Summer－Winter,Horse－Zebra,the SIFID value of the generated imageis reduced by 30% . The proposed method can better preserve
the content information of the image while generating detailed and realistichigh-quality images.