计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (3): 231-238.DOI: 10.3778/j.issn.1002-8331.1911-0397

• 图形图像处理 • 上一篇    下一篇

多判别器循环生成对抗网络的素描人脸合成

周华强,曹林,杜康宁   

  1. 1.北京信息科技大学 光电测试技术及仪器教育部重点实验室,北京 100101
    2.北京信息科技大学 信息与通信工程学院,北京100101
  • 出版日期:2021-02-01 发布日期:2021-01-29

Sketch Face Synthesis Based on Multi-discriminator Cyclic Generative Adversarial Network

ZHOU Huaqiang, CAO Lin, DU Kangning   

  1. 1.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument, Beijing Information Science and Technology University, Beijing 100101, China
    2.School of Information and Communication Engineering, Beijing Information Science and Technology University, Beijing 100101, China
  • Online:2021-02-01 Published:2021-01-29

摘要:

素描人脸合成在娱乐和刑侦领域具有重要应用价值。为了解决传统素描人脸合成方法生成图像面部细节模糊,缺失真实感等问题,改进了CycleGAN网络结构,提出一种基于多判别器循环生成对抗网络的素描人脸合成方法。该方法选取残差网络作为生成网络模型,在生成器隐藏层中增加多个判别器,提高网络对生成图像细节特征的提取能力;并建立了重构误差约束映射关系,最小化生成图像与目标图像之间的距离。通过在CUHK和AR人脸数据库中的对比实验,证明了相比于原始CycleGAN框架该方法性能有明显提升;相比于目前领先的方法,所提方法生成的素描图像细节特征更清晰,真实感更强。

关键词: 素描人脸合成, 生成对抗网络, 残差网络, 多判别器网络, 深度学习

Abstract:

Sketch face synthesis has important application value in the field of entertainment and criminal investigation. In order to solve the problem of fuzzy face details and lack of realistic sense in the traditional sketch face synthesis method, the CycleGAN framework is improved and a sketch face synthesis method based on multiple discriminator cycle generation antagonistic network is proposed. The residual network is used in proposed method as generating network model, and multiple discriminator are added in the hidden layer of generator to improve the performance of network to extract the detailed features of the generated image. The reconstructed error constrained mapping relation is established to minimize the distance between the generated image and the target image. Through the comparison experiment of CUHK and AR face database, it is proved that the performance of this method is obviously improved compared with that of the original CycleGAN framework. Compared with the current leading methods, the detailed features of the sketch images generated by the method proposed in this paper are clearer and more realistic.

Key words: face sketch synthesis, generative adversarial network, residual network, multi-discriminator networks, deep learning