[1]郭美钦,江健民.人脸图像风格迁移的改进算法[J].深圳大学学报理工版,2019,(No.3(221-346)):230-236.[doi:10.3724/SP.J.1249.2019.03230]
 GUO Meiqin and JIANG Jianmin.Spatially-robust image style transfer for headshot portraits[J].Journal of Shenzhen University Science and Engineering,2019,(No.3(221-346)):230-236.[doi:10.3724/SP.J.1249.2019.03230]
点击复制

人脸图像风格迁移的改进算法()
分享到:

《深圳大学学报理工版》[ISSN:1000-2618/CN:44-1401/N]

卷:
期数:
2019年No.3(221-346)
页码:
230-236
栏目:
【电子与信息科学】
出版日期:
2019-05-20

文章信息/Info

Title:
Spatially-robust image style transfer for headshot portraits
作者:
郭美钦江健民
深圳大学计算机与软件学院,广东深圳 518060
Author(s):
GUO Meiqin and JIANG Jianmin
College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060,Guangdong Province, P.R.China
关键词:
计算机图像处理计算机神经网络图像风格迁移卷积神经网络归一化互相关仿射变换语义分割人脸图像
Keywords:
computer image processing convolutional neural network image style transfer convolutional neural network normalizedcross-correlation affine transformation semantic segmentation headshot portioit
分类号:
TP 391
DOI:
10.3724/SP.J.1249.2019.03230
文献标志码:
A
摘要:
图像风格迁移是对给定的输入图像,采用计算机手段自动将任意其他图像的风格迁移到输入图像中,使两者风格完全相同,但保持原图像内容不变.本研究通过实验发现,当这种风格迁移用到人脸图像时,由于嘴唇、眼睛等部位对空间位置差异比较敏感,容易导致风格迁移效果失真.基于(deep photo style transfer,DPST)算法,提出对空间差异鲁棒的人脸图像风格迁移的改进算法,通过最大化输入图像和风格图像各部分语义的分割内容在深度卷积神经网络中特征图矩阵的归一化互相关,对风格图像进行对应的仿射变换,将变换后的一簇图像作为风格参考来完成风格迁移,有效减轻了输入图像和风格图像各部分语义分割内容在空间上存在较大差异时产生的重影.选取空间差异较大的输入图像与风格图像与DPST算法进行对比,实验结果表明,当两张图像各语义分割内容在空间位置相差较大时,新算法仍能产生较好的效果,鲁棒性更佳.
Abstract:
Image style transfer addresses such a research problem that, given an input image, we can transfer the style of any image into the input, and hence change the input style accordingly without modifying its original content. Our empirical studies discover that, when the existing style transfer techniques are applied to headshot portraits, noticeable distortions could be incurred due to the fact that the relative location of those facial components, such as lips, noses, and eyes, are sensitive to spatial differences. Based on the (deep photo style transfer, DPST) method reported by Luan et al, a spatially-robust image style transfer algorithm for headshot portraits is proposed in this paper to improve the robustness of image style transfer. The proposed algorithm provides an improved solution for the problem that, when segmented regions between the input image and the style image are significantly different, the transferred content could incur ghost shadows. By maximizing the normalized cross-correlation of the feature matrix in the trained convolutional neural network, we propose to apply affine transforms to the style image before it is used for the style transfer. In this way, the transferred style images can be used as style references for all segmented regions, providing advantages that the content to be style transferred is made adaptive to the difference between the input and the style reference and hence the incurred ghost shadow can be minimized or eliminated. Experiments show that our proposed can produce good results even when the semantic segmentation of the two images have large spatial differences, and the proposed also displays better robustness than the DPST(Deep Photo Style Transfer) algorithm.

相似文献/References:

[1]殷夫,纪震,周家锐,等.PET/CT图像质量主观评价与感知模型[J].深圳大学学报理工版,2015,32(No.2(111-220)):205.[doi:10.3724/SP.J.1249.2015.02205]
 Yin Fu,Ji Zhen,Zhou Jiarui,et al.Subjective assessment and perception model of PET/CT image quality[J].Journal of Shenzhen University Science and Engineering,2015,32(No.3(221-346)):205.[doi:10.3724/SP.J.1249.2015.02205]
[2]陈玉佳,姜波.基于小波神经网络的加工番茄产量预测模型[J].深圳大学学报理工版,2015,32(No.5(441-550)):546.[doi:10.3724/SP.J.1249.2015.05546]
 Chen Yujia and Jiang Bo.A wavelet neural network model for processing tomato yield forecasting[J].Journal of Shenzhen University Science and Engineering,2015,32(No.3(221-346)):546.[doi:10.3724/SP.J.1249.2015.05546]

更新日期/Last Update: 2019-04-22