[1]李学相,曹淇,刘成明.基于无配对生成对抗网络的图像超分辨率重建[J].郑州大学学报(工学版),2021,42(05):1-6.[doi:10.13705/j.issn.1671-6833.2021.05.018]
　LI Xuexiang,CAO Qi,LIU Chengming.Image Super-resolution Based on No Match Generative Adversarial Network[J].Journal of Zhengzhou University (Engineering Science),2021,42(05):1-6.[doi:10.13705/j.issn.1671-6833.2021.05.018]

点击复制

基于无配对生成对抗网络的图像超分辨率重建()

分享到：

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:

42

期数:

2021年05期

页码:

1-6

栏目:

出版日期:

2021-09-10

文章信息/Info

Title:

Image Super-resolution Based on No Match Generative Adversarial Network

作者:

李学相; 曹淇; 刘成明

郑州大学软件学院;

Author(s):

LI Xuexiang;  CAO Qi;  LIU Chengming

School of Software, Zhengzhou University;

关键词:

超分辨率; 深度学习; 生成对抗网络; 无配对; 二阶统计量

Keywords:

super-resolution;  deep learning;  generative adversarial network;  no matching;  second-order statistic

DOI:

10.13705/j.issn.1671-6833.2021.05.018

文献标志码:

A

摘要:

图像超分辨率重建一直是计算机图像处理领域的热门课题。近来，基于生成对抗网络的图像超分辨率重建方法能够感知高频的纹理细节，取得了良好的重建效果。但是，生成对抗网络图像重建在图像质量方面表现出高度的不稳定性。对此，我们在SRGAN基础上提出了一个新的基于无配对图像的模型NM-SRGAN。首先，我们利用循环生成对抗网络（Cycle-gan）作为预处理模块，用于训练无配对数据集并获得一个更好的输入图像，在保留残差块基础结构的同时删除了残差块中的BN层，解决了结果不稳定的问题。同时，我们基于二阶统计量比一阶统计量更能捕捉区域信息的原理，使用协方差矩阵捕捉图像的二阶信息，在感知损失的基础上增加了二阶损失函数，使模型更加注重于捕捉图像细节区域部分的变化。最后，我们将感知损失中的VGG网络的损失函数改为一阶梯度的VGG损失函数，专注于提升图像的边缘纹理细节。我们对提出的NM-SRGAN在4个标准数据集上进行测试评估并将部分结果图进行细节部分比较，使用客观评价标准峰值信噪比与结构化相似化度量对结果图进行测试，并与经典卷积方法及SRGAN进行比较。实验结果表明，我们的方法在稳定性及图像质量、细节方面较SRGAN及经典算法均有较好的改善。

Abstract:

Image super-resolution reconstruction based on generative adversarial networks (GAN) is subject to the dataset training with an unstable result. To solve this problem, a new NM-SRGAN model is established. The cycle-gan is firstly used as the preprocess module to make the model free from the dataset for training with better input of the image, and the model cancels BN layer to solve the unstable results. Besides, covariance matrix is adopted to capture the second-order information of the image, and second-order loss function is added with a focus on the changes of the image details. The new VGG loss function is used to improve the marginal texture of the image. The proposed NM-SRGAN model is verified by four standard datasets, and the resulting images are assessed by the objective evaluation indices. Compared with the existing models, NM-SRGAN model has an improved evaluation value of 0.19, 0.03, 0.13, and 0.02 dB, respectively, reaching up to the maximum among the four datasets. Results show that the proposed method, compared with traditional algorithms, has achieved better improvements in stability and image quality with better details.

常用功能