Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (9): 1612-1620.DOI: 10.3778/j.issn.1673-9418.1910067

Previous Articles    

Single Image Super-Resolution Reconstruction Method for Generative Adversarial Network

PENG Yanfei, GAO Yi, DU Tingting, SANG Yu, ZI Lingling   

  1. School of Electronic and Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2020-09-01 Published:2020-09-07

生成对抗网络的单图像超分辨率重建方法

彭晏飞高艺杜婷婷桑雨訾玲玲   

  1. 辽宁工程技术大学 电子与信息工程学院,辽宁 葫芦岛 125105

Abstract:

The super-resolution reconstruction method based on deep convolutional neural network has a high peak signal-to-noise ratio (PSNR), but the reconstruction results have the problem of lack of high-frequency information and texture details and poor visual perception under large-scale factors. Aiming at this problem, a single image super-resolution reconstruction method based on generative adversarial network is proposed. Firstly, the hinge loss in the migration support vector machine is taken as the objective function, and then the Charbonnier loss which is more stable and more anti-noise is used instead of the L2 loss function. Finally, the batch normalization layer which is unfavorable to the super resolution of the image in the residual block and discriminator is removed, and the spectral normalization is used in the generator and discriminator to reduce the computational overhead and stabilize the model training. The experimental results of 4X upscaling show that compared with other comparison methods, the PSNR value of the reconstructed image is improved by up to 4.6 dB and the SSIM value is increased by 0.1, and the test time is shorter. The experimental data and effect diagram show that the super-resolution image reconstructed by this method has better visual effect and higher PSNR and SSIM values.

Key words: super-resolution reconstruction, generative adversarial network (GAN), deep learning, convolutional neural network (CNN), loss function

摘要:

基于深度卷积神经网络的超分辨率重建方法虽然有较高的峰值信噪比(PSNR),但重建结果在大尺度因子下存在缺乏高频信息和纹理细节,视觉感知效果差的问题。针对这一问题,提出了一种基于生成对抗网络的单图像超分辨率重建方法。首先迁移支持向量机中的hinge损失作为目标函数,其次使用更加稳定、抗噪性更强的Charbonnier损失代替L2损失函数,最后去掉了残差块和判别器中对图像超分辨率不利的批规范化层,并在生成器和判别器中使用谱归一化来减小计算开销,稳定模型训练。实验结果表明,在4倍放大尺度因子下,相较其他对比方法,该方法重建图像的PSNR值最高提升4.6 dB,SSIM值最高提升0.1,测试时间较短。实验数据和效果图均表明该方法重建的超分辨率图像视觉效果较好,且有更高的PSNR和SSIM值。

关键词: 超分辨率重建, 生成对抗网络(GAN), 深度学习, 卷积神经网络(CNN), 损失函数