作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (7): 251-259. doi: 10.19678/j.issn.1000-3428.0055551

• 图形图像处理 • 上一篇    下一篇

基于多尺度反向投影的图像超分辨率重建算法

熊亚辉a,b, 陈东方a,b, 王晓峰a,b   

  1. 武汉科技大学 a. 计算机科学与技术学院;b. 湖北省智能信息处理与实时工业系统重点实验室, 武汉 430065
  • 收稿日期:2019-07-22 修回日期:2019-09-18 发布日期:2019-09-25
  • 作者简介:熊亚辉(1994-),男,硕士研究生,主研方向为图像处理、超分辨率重建;陈东方,教授、博士;王晓峰,副教授、博士。
  • 基金资助:
    国家自然科学基金(61572381,61273225)。

Super-Resolution Image Reconstruction Algorithm Based on Multi-Scale Back Projection

XIONG Yahuia,b, CHEN Dongfanga,b, WANG Xiaofenga,b   

  1. a. School of Computer Science and Technology;b. Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System, Wuhan University of Science and Technology, Wuhan 430065, China
  • Received:2019-07-22 Revised:2019-09-18 Published:2019-09-25

摘要: 为解决当前主流图像超分辨率重建算法对低分辨率图像中细节信息利用不够充分的问题,提出一种基于多尺度反向投影的图像超分辨率重建算法。使用多个不同尺度的卷积核从浅层特征提取层中提取出不同维度的特征信息,输入到反向投影模块后,交替使用升采样和降采样来优化高分辨率和低分辨率图像的投影误差,同时运用残差学习的思想将升采样和降采样阶段提取到的特征使用级联的方式进行连接,从而提升图像的重建效果。实验结果表明,在Set5、Set14和Urban100数据集上,与Bicubic、SRCNN、ESPCN、VDSR和LapSRN 5种主流算法相比,该算法的峰值信噪比和结构相似性均有所提高。

关键词: 图像超分辨率重建算法, 多尺度, 反向投影, 迭代式升采样和降采样, 深度学习

Abstract: In order to solve the problem that existing mainstream super-resolution image reconstruction algorithms fail to fully utilize the detailed information in Low-Resolution(LR) images,this paper proposes a super-resolution image reconstruction algorithm based on multi-scale back projection.The algorithm uses multiple convolutional kernels of different scales to extract feature information of different dimensions from the shallow feature extraction layer.Then the extracted feature information is input into the back projection module,and the upsampling and downsampling methods are used alternatively to optimize the projection error of High-Resolution(HR) and LR images.Also,the idea of residual learning is used to connect the features extracted in the upsampling and downsampling stages in a cascade manner,so as to improve the image reconstruction effect.Experimental results on the Set5,Set14 and Urban100 datasets show that the proposed algorithm improves the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity(SSIM) compared with the five mainstream algorithms such as Bicubic,SRCNN,ESPCN,VDSR and LapSRN.

Key words: super-resolution image reconstruction algorithm, multi-scale, back projection, iterative upsampling and downsampling, deep learning

中图分类号: