计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (12): 201-208.DOI: 10.3778/j.issn.1002-8331.1905-0391

• 图形图像处理 • 上一篇    下一篇

基于残差网络的快速图像风格迁移研究

薛楠,段锦,张兵,王晓宇,于林韬   

  1. 长春理工大学 电子信息工程学院,长春 130000
  • 出版日期:2020-06-15 发布日期:2020-06-09

Research on Fast Image Style Transformation Based on Residual Network

XUE Nan, DUAN Jin, ZHANG Bing, WANG Xiaoyu, YU Lintao   

  1. School of Electronic Information Engineering, Changchun University of Science and Technology, Changchun 130000, China
  • Online:2020-06-15 Published:2020-06-09

摘要:

图像风格迁移技术是指将一幅图像通过学习(利用卷积神经网络)名画风格,转换为与名画风格相近的图像。Gatys提出的NAAS利用VGG网络设计了一个损失网络,通过反复迭代得到风格迁移图像。Li Feifei在NAAS的基础上引入残差网络,利用残差元的快捷连接特性加速计算。主要针对以下两个方面提出了改进:对经典残差元结构进行调整,将标准卷积转换为点卷积和深度卷积,在保证卷积效果的同时降低计算量;对损失网络进行简化,该模型中第四、第五层在结构上高度一致,并且这两层的风格还原与内容重建效果基本相同,因此删去第五层并相应调整结构参数,去掉冗余参数,在降低参数量的同时保证风格还原与内容重建的效果。

关键词: 图像风格迁移, 残差网络, VGG网络

Abstract:

Image style migration technology refers to the conversion of an image into a similar image to a famous painting style through learning (using a convolutional neural network). The NAAS proposed by Gatys uses the VGG network to design a loss network, and the style migration image is obtained through repeated iterations. Li Feifei has introduced the residual network on the basis of NAAS, and the calculation is accelerated by using the fast connection feature of the residual element. This paper proposes improvements for the following two aspects:the classical residual element structure is adjusted, and the standard convolution is converted into point convolution and deep convolution, which reduces the amount of calculation while ensuring the convolution effect. The loss network is simplified. The fourth and fifth layers in the model are highly consistent in structure, and the style restoration of the two layers is basically the same as the content reconstruction effect. Therefore, the fifth layer is deleted and the structural parameters are adjusted accordingly, and the redundant parameters are removed. The effect of style restoration and content reconstruction is guaranteed while reducing the amount of parameters.

Key words: image style transformation, residual network, VGG network