双路多尺度残差网络的图像超分辨率重建

全文下载： 202103017.pdf

文章编号： 1672-6987（2021）03-0107-06； DOI： 10.16351/j.1672-6987.2021.03.017

胡锐, 赵佰亭*, 贾晓芬（安徽理工大学 电气与信息工程学院，安徽 淮南 232001）

摘要： 针对传统单幅图像超分辨率重建方法出现的边缘特征模糊问题，提出了一种双路多尺度残差网络（BMRN）的重建方法。首先直接对低分辨率图像进行特征提取，较大程度保留特征信息；再构建多个独立的双路多尺度残差网络提取高频信息，其中残差连接的引入可以有效解决网络加深导致的梯度消失问题，双路多尺度结构可以相互补充卷积中的尺度信息，改善网络中的信息流；最后通过上采样操作，得到重建高分辨率图像。在数据集Set5上进行的×3尺度的重建结果表明：与Bicubic、SRCNN和VDSR等传统方法相比，BMRN的平均PSNR/SSIM分别提高了337 dB/0053 2、101 dB/0012 4和009 dB/0000 4。实验数据充分说明：BMRN能够较好的恢复图像轮廓特征，获得了较高的PSNR、SSIM和更好的视觉效果。

关键词： 图像超分辨率重建； 多尺度卷积； 残差网络； 亚像素卷积

中图分类号： TP 391文献标志码： A

引用格式： 胡锐, 赵佰亭, 贾晓芬. 双路多尺度残差网络的图像超分辨率重建［J］. 青岛科技大学学报(自然科学版), 2021, 42(3): 107-112.

HU Rui， ZHAO Baiting， JIA Xiaofen. Image super-resolution reconstruction based on binary channels multi-scale residual network［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2021， 42（3）： 107-112.

Image Super-resolution Reconstruction Based on Binary Channels

Multi-scale Residual Network

HU Rui， ZHAO Baiting， JIA Xiaofen

(School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan 232001, China)

Abstract: In order to solve the problem of fuzzy edge features in traditional single image super-resolution reconstruction method, a binary channel multi-scale residual network (BMRN) reconstruction method is proposed. Firstly, the low-resolution image features are extracted directly to retain the feature information to a large extent. Secondly, several independent dual channel multi-scale residual networks are constructed to extract high-frequency information. The introduction of residual connection can effectively solve the problem of gradient disappearance caused by network deepening, and the binary channel multi-scale structure can complement the scale information in convolution to improve the information flow in the network; finally, the multi-scale residual network can be used to extract the high-frequency information through the up sampling operation, the reconstructed high-resolution image is obtained. The 3-scale reconstruction results on Set5 show that the average PSNR/SSIM of BMRN is 337 dB/00532, 101 dB/00124 and 009 dB/00004 respectively compared with traditional algorithms such as Bicubic, SRCNN and VDSR. The experimental data fully show that BMRN can restore the image contour features, and obtain higher PSNR, SSIM and better visual effect.

Key words: image super-resolution reconstruction; multi-scale convolution; residual network; sub-pixel convolution

收稿日期： 2020-07-04

基金项目： 安徽省高校自然科学研究重大资助项目（KJ2018ZD008）；国家自然科学基金项目（61501006）；国家重点研发计划专项资助项目（2016YFC0600908）；安徽省重大科技专项项目（16030901012）.

作者简介： 胡锐（1995—），男，硕士研究生.*通信联系人.