计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (4): 680-687.DOI: 10.3778/j.issn.1673-9418.1905082

• 图形图像 • 上一篇    下一篇

生成式对抗网络在超分辨率图像重建中的应用

汪鑫耘,李丹   

  1. 安徽工业大学 电气与信息工程学院,安徽 马鞍山 243032
  • 出版日期:2020-04-01 发布日期:2020-04-10

Application of Generative Adversarial Network in Super-Resolution Image Reconstruction

WANG Xinyun, LI Dan   

  1. College of Electrical and Information Engineering, Anhui University of Technology, Ma'anshan, Anhui 243032, China
  • Online:2020-04-01 Published:2020-04-10

摘要:

针对现有卷积神经网络图像超分辨率算法容易出现过拟合、损失函数的收敛性不足等问题,结合超分辨率算法和生成式对抗网络(GAN)理论,设计一种基于生成式对抗网络的超分辨率算法PESRGAN用于恢复四倍下采样的图像。首先使用残差密集块(RDB)作为基本结构单元,有效避免了过拟合问题;其次使用双层特征损失并使用渗透指数(PI)作为损失的权值,更好地去学习低分辨率到高分辨率图像之间的映射关系;同时使用VGG19作为判别网络高分辨率图像进行分类;最后使用经典数据集,将PESRGAN算法与双三次插值(Bicubic)、SRGAN、ESRGAN算法在客观参数和主观视觉效果进行对比。实验结果表明:在经典数据集上,PESRGAN的平均峰值信噪比(PSNR)达到25.4 dB、平均结构相似性(SSIM)达到0.73,平均渗透指数(PI)达到1.15,在客观参数和主观评价上均优于其他算法,证明了PESRGAN有良好的超分辨率重建的效果。

关键词: 卷积神经网络(CNN), 超分辨率图像重建, 生成式对抗网络(GAN), 四倍采样

Abstract:

Aiming at the problems of over-fitting of existing convolutional neural network image super-resolution algorithm and insufficient convergence of loss function, combined with super-resolution algorithm and generative adversarial network (GAN) theory, the super-resolution algorithm PESRGAN (permeability enhanced super-resolution generative adversarial networks) is used to recover quadruple downsampled images. Firstly, the residual dense block (RDB) is used as the basic structural unit, which effectively avoids the over-fitting problem. Secondly, the double-layer feature loss is used and the permeability index (PI) is used as the weight of the loss, so that it can better learn the mapping relationship between low-resolution and high-resolution images; VGG19 is used as the discriminant network high-resolution image for classification. Finally, PESRGAN algorithm is compared with Bicubic, SRGAN(super-resolution using a generative adversarial network) and ESRGAN (enhanced super-resolution generative adver-sarial networks) algorithm in objective parameters and subjective visual effects using classical data sets. The experi-mental results show that the average peak signal to noise ratio (PSNR) of PESRGAN is 25.4 dB, the average stru-ctural similarity index (SSIM) is 0.73 and the average PI is 1.15 on the classical data sets. The objective parameters and subjective evaluation of PESRGAN are superior to other algorithms, which proves that PESRGAN has good super-resolution reconstruction effect.

Key words: convolutional neural network (CNN), super-resolution image reconstruction, generative adversarial network (GAN), quadruple sampling