Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (11): 219-223.DOI: 10.3778/j.issn.1002-8331.2002-0370

Previous Articles     Next Articles

Research on Face Super-Resolution Reconstruction Algorithm Based on Generative Adversarial Networks

JIANG Wenjie, LUO Xiaoshu, DAI Qinxuan   

  1. College of Electronic Engineering, Guangxi Normal University, Guilin, Guangxi 541004, China
  • Online:2021-06-01 Published:2021-05-31

基于对抗网络人脸超分辨率重建算法研究

蒋文杰,罗晓曙,戴沁璇   

  1. 广西师范大学 电子工程学院,广西 桂林 541004

Abstract:

Due to the influence of illumination and imaging equipment and other factors, the single frame of face image collected has a low resolution, which makes it impossible to perform accurate face recognition. Therefore, image super-resolution reconstruction is required. However, in the process of face super-resolution reconstruction using SRGAN model, gradient disappearance or explosion is easy to occur, which seriously affects the accuracy and quality of the reconstructed image. According to the above problem, this paper puts forward the improvement based on the generated against network face super-resolution reconstruction algorithm, on the basis of SRGAN combination of WGA-N introduction out divergence, and to maximize its [T] get optimal scalar functions, implementation constraints can remove Lipschit-z Wassertein distance can be obtained directly, and generation network of the objective function is obtained by minimizing Wassertein distance, finally the improved model can improve the quality of reconstruction image. The experimental results show that this method can generate high-resolution face images and is better than DRCN, FSRCNN, SRGAN_WGAN, VDSR and DRRN models in both subjective and objective evaluation indexes.

Key words: face super-resolution, generative adversarial network, Wasserstein distance, Wasserstein divergence

摘要:

由于受到光照和成像设备等条件因素的影响,采集到的单帧人脸图像分辨率低,无法进行准确人脸识别,所以需要图像超分辨率重建。而利用SRGAN模型在进行人脸超分辨率重建过程中,易出现梯度消失或爆炸的问题,严重影响了重建图像的精度和质量。针对上述问题,提出了基于生成对抗网络的改进人脸超分辨率重建算法,在SRGAN结合WGA-N的基础上引入Wasserstein散度,并将其最大化得到最优化标量函数[T],实现去掉Lipschit-z约束能够直接得到Wassertein距离,并通过最小化Wassertein距离得到生成网络的目标函数,最终改进的模型提高了重建图像的质量。实验结果表明,该方法能够生成高分辨率的人脸图像,在主观视觉和客观评价指标均同比优于DRCN、FSRCNN、SRGAN_WGAN、VDSR和DRRN模型。

关键词: 人脸超分辨率重建, 生成对抗网络, Wasserstein距离, Wasserstein散度