赵小强^{1, 2, 3,,},
宋昭漾¹
1.兰州理工大学电气工程与信息工程学院 ??兰州 ??730050
2.甘肃省工业过程先进控制重点实验室 ??兰州 ??730050
3.兰州理工大学国家级电气与控制工程实验教学中心 ??兰州 ??730050
基金项目:国家科学自然基金(61763029, 61873116)

详细信息

作者简介:赵小强：男，1969年生，博士生导师，教授，主要研究方向为故障诊断，图像处理，生产调度等
宋昭漾：男，1995年生，硕士生，研究方向为图像处理

通讯作者:赵小强　xqzhao@lut.cn

中图分类号:TP391

计量

文章访问数:2071
HTML全文浏览量:907
PDF下载量:72
被引次数:0

出版历程

收稿日期:2019-01-15
修回日期:2019-06-30
网络出版日期:2019-07-19
刊出日期:2019-10-01

Super-Resolution Reconstruction of Deep Residual Network with Multi-Level Skip Connections

Xiaoqiang ZHAO^{1, 2, 3,,},
Zhaoyang SONG¹
1. College of Electrical Engineering and Information Engineering, Lanzhou University of Technology, Lanzhou 730050, China
2. Key Laboratory of Gansu Advanced Control for Industrial Processes, Lanzhou University of Technology, Lanzhou 730050, China
3. National Experimental Teaching Center of Electrical and Control Engineering, Lanzhou University of Technology, Lanzhou 730050, China
Funds:The National Natural Science Foundation of China (61763029, 61873116)


摘要
摘要:由于快速的卷积神经网络超分辨率重建算法(FSRCNN)卷积层数少、相邻卷积层的特征信息之间缺乏关联性，因此难以提取到图像深层信息导致图像超分辨率重建效果不佳。针对此问题，该文提出多级跳线连接的深度残差网络超分辨率重建方法。首先，该方法设计了多级跳线连接的残差块，在多级跳线连接的残差块基础上构造了多级跳线连接的深度残差网络，解决相邻卷积层的特性信息缺乏关联性的问题；然后，使用随机梯度下降法(SGD)以可调节的学习率策略对多级跳线连接的深度残差网络进行训练，得到该网络超分辨率重建模型；最后，将低分辨率图像输入到多级跳线连接的深度残差网络超分辨率重建模型中，通过多级跳线连接的残差块得到预测的残差特征值，再将残差图像和低分辨率图像组合在一起转化为高分辨率图像。该文方法与bicubic, A+, SRCNN, FSRCNN和ESPCN算法在Set5和Set14测试集上进行了对比测试，在视觉效果和评价指标数值上该方法都优于其它对比算法。
关键词:超分辨率重建/
深度残差网络/
多级跳线连接的残差块/
随机梯度下降法
Abstract:The Fast Super-Resolution Convolutional Neural Network algorithm (FSRCNN) is difficult to extract deep image information due to the small number of convolution layers and the correlation lack between the feature information of adjacent convolutional layers. To solve this problem, a deep residual network super-resolution reconstruction method with multi-level skip connections is proposed. Firstly, a residual block with multi-level skip connections is designed to solve the problem that the characteristic information of adjacent convolutional layers lacks relevance. A deep residual network with multi-level skip connections is constructed on the basis of the residual block. Then, the deep residual network connected to the multi-level skip is trained by using the adaptive gradient rate strategy of Stochastic Gradient Descent (SGD) method and the network super-resolution reconstruction model is obtained. Finally, the low-resolution image is input into the deep residual network super-resolution reconstruction model with the multi-level skip connections, and the residual eigenvalue is obtained by the residual block connected the multi-level skip connections. The residual eigenvalue and the low resolution image are combined and converted into a high resolution image. The proposed method is compared with the bicubic, A+, SRCNN, FSRCNN and ESPCN algorithms in the Set5 and Set14 test sets. The proposed method is superior to other comparison algorithms in terms of visual effects and evaluation index values.
Key words:Super-resolution reconstruction/
Deep residual network/
Residual network with multi-level skip connections/
Stochastic Gradient Descent (SGD)



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=268d240f-ac5b-4129-92f7-cd9e901db4fd