«上一篇
 文章快速检索 高级检索

 应用科技  2020, Vol. 47 Issue (4): 8-13  DOI: 10.11991/yykj.201912020 0

### 引用本文

SHI Yongxiang, JIANG Bin, HUANG Yongzhuo, et al. Infrared image super-resolution reconstruction based on deep learning[J]. Applied Science and Technology, 2020, 47(4): 8-13. DOI: 10.11991/yykj.201912020.

### 文章历史

1. 国家电网溧阳市供电公司，江苏 溧阳 213300;
2. 河海大学 物联网工程学院，江苏 常州 213000

Infrared image super-resolution reconstruction based on deep learning
SHI Yongxiang1, JIANG Bin1, HUANG Yongzhuo1, YANG Guisheng1, LI Qingwu2, ZHANG Zhiliang2
1. State Grid Liyang Power Supply Company, Liyang 213300, China;
2. School of Internet of Things Engineering, Hohai University, Changzhou 213000, China
Abstract: In order to improve the resolution of infrared image, this paper constructs an IEDSR (enhanced deep residual networks for infrared image super-resolution) network for infrared image super-resolution reconstruction. Based on the EDSR (enhanced deep residual networks for single image super-sesolution) network model, a pooling layer is added to the network, which avoids the problem that removing BN (batch normalization) layer from EDSR network may bring training difficulty. At the same time, considering the low contrast of infrared image and the characteristics of not obvious texture, a new convolution layer and activation layer are added to the residual block, which is helpful to recover the local details of the image by increasing the depth of network and expanding the receptive field of the local residual module. Finally, we use the enhanced prediction algorithm to optimize the reconstructed image and improve the reconstruction accuracy. The experimental results show that the infrared image reconstructed by this algorithm has better subjective visual effect and objective index than traditional infrared image reconstruction method, and has higher practical value.
Keywords: neural network    deep learning    residual network    infrared image    super-resolution reconstruction    pool layer    receptive field    enhanced forecasting

1 IEDSR网络结构

IEDSR网络以残差网络结构作为骨干网络，和EDSR网络相似，模型只需要学习图像高频信息，提高了学习速度；同时在网络输出端都使用了反卷积层，实现图像空间分辨率的提升。不同的是，IEDSR网络针对红外图像分辨率低、信噪比差和对比度低等特性，对残差模块进行了改进：加入了新的卷积层和池化层，扩大感受野，提高模型的学习能力。另外，在模型预测时使用了增强预测算法来提高精准度。

1.1 残差网络

1.2 残差块的改进

 $F(x) = \max \{ 0,U(x) + x\}$

 $R = (R' - 1) \times S + K$ (1)

1.3 增强预测算法

2 实验过程与结果分析

2.1 训练过程

 ${\rm{Loss}} = {\rm{MSE}} = \frac{1}{{MN}}\sum\limits_{i = 0}^{M - 1} {\sum\limits_{j = 0}^{N - 1} {{{\left[ {{{Y}}(i,j) - \hat {{Y}}(i,j)} \right]}^2}} }$ (2)

2.2 训练策略

2.3 实验结果分析与比较

 ${\rm{PSNR}} = 10\lg \Bigg(\frac{{{\rm{MA}}{{\rm{X}}^2}}}{{{\rm{MSE}}}}\Bigg)$

 ${\rm{SSIM}}({{Y}},\hat {{Y}}) = \frac{{(2{\mu _Y}{\mu _{\hat Y}} + C)(2{\sigma _{Y\hat Y}} + C')}}{{(\mu _Y^2 + \mu _{\hat Y}^2 + C)(\sigma _Y^2 + \sigma _{\hat Y}^2 + C')}}$