欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 视觉与图像 • 上一篇    下一篇

基于字典学习的 HDR 照片风格转移方法

  

  1. 1. 上海大学上海电影学院，上海 200072；2. 上海交通大学计算机科学与工程系，上海 200042
  • 出版日期:2017-10-31 发布日期:2017-11-03
  • 基金资助:
    国家自然科学基金项目(61303093，61402278)；上海市科委攻关项目(16511101300)；上海工程技术研究中心建设专项(16dz2251300)；上
    海市自然科学基金项目(14ZR1415800)；上海市教委科研创新项目(14YZ023)；上海大学电影学高峰学科项目

The Style Transfer of HDR Image Based on Dictionary Learning

  1. 1. Shanghai Film Academy, Shanghai University, Shanghai 200072, China;
    2. Department of Computer Science and Technology, Shanghai Jiao Tong University, Shanghai 200042, China
  • Online:2017-10-31 Published:2017-11-03

PDF (PC)

216

摘要： HDR 照片风格，包括自然、浮雕、绘画、油画等，能够展现丰富的艺术特质，具
有风格显著、细节丰富、颜色饱满等特点。但是，传统的 HDR 风格生成方法交互繁琐、费时费
力，并且生成的 HDR 效果稳定性差。因此，为了简化交互、提升效果，提出了一种 HDR 照片风
格转移方法，通过给定一张 HDR 参考照片，借助颜色转移和字典学习技术，将 HDR 风格特征转
移到源照片上，从而自动生成 HDR 照片效果。首先，借助梯度保持的颜色转移技术，将 HDR 参
考照片的颜色特征转移到源照片上；然后，对 HDR 参考照片提取细节特征，利用 K_SVD 算法
进行字典训练，形成细节的过完备字典集；接着，利用细节字典集对源照片进行稀疏重建，生成
与参考照片特征一致的细节；最后，将颜色转移结果和细节重建结果合并，生成 HDR 风格的新
照片。实验表明该方法应用到多种 HDR 风格，都可以获得跟参考照片风格一致的视觉效果。

关键词: HDR 风格, 颜色转移, 字典学习, K-SVD, 稀疏重建

Abstract: The style of HDR image includes natural, reliefs, drawings, paintings and so on. It shows the
rich artistic qualities, has the characteristics of remarkable style, rich in detail, full color and others.
However, the conventional method for generating HDR style is tedious, time consuming and the effect
of HDR is instability. Therefore, in order to simplify the interaction and enhance the effect of the result,
we proposed a style transfer method for HDR image by giving an HDR reference image and using color
transfer and dictionary learning technology to transfer the features of HDR photograph to the source
image so that to generate the HDR image automatically. First, with the help of gradient-preserving color
transfer technology, the color feature of the HDR reference image was transferred to the source photo.
Then, extracting the detail features of the HDR reference image and using K_SVD algorithm to train a
dictionary to form an over-complete dictionary set of the details. Next, extracting the detail features of
source image and using the dictionary set to sparse reconstruction and then generating the right HDR
image with detail features. Finally, combining the result of color transfer with the sparse reconstruction
result of the source photo to generate a new style of HDR image. The experiment results show that the
method applied to a variety of HDR styles can obtain the same visual effects as the reference image.

Key words: HDR style, color transfer, dictionary learning, K-SVD, sparse reconstruction