图像超分辨率全局残差递归网络

计算机科学 ›› 2019, Vol. 46 ›› Issue (6A): 230-233.

图像超分辨率全局残差递归网络

张雷¹, 胡博文¹, 张宁², 王茂森²

沈阳航空航天大学电子信息工程学院沈阳110136¹;
上海航天电子技术研究所上海201109²

出版日期:2019-06-14 发布日期:2019-07-02
通讯作者: 胡博文(1992-),男,硕士生,主要研究方向为图像超分辨率、图像压缩等,E-mail:767464760@qq.com
作者简介:张雷(1972-),男,博士,副教授,主要研究方向为图像压缩、图像识别等,E-mail:rd_zhangl@126.com;张宁(1982-),男,博士,高级工程师,主要研究方向为卫星数据智能处理、空间电子系统等;王茂森(1985-),男,硕士,工程师,主要研究方向为卫星计算机等。
基金资助:
本文受国家自然科学基金(61671037),上海航天科技创新基金(SAST2016090)资助。

Global Residual Recursive Network for Image Super-resolution

ZHANG Lei¹, HU Bo-wen¹, ZHANG Ning², WANG Mao-sen²

College of Electronic Information Engineering,Shenyang Aerospace University,Shenyang 110136,China¹;
Space Electronic Technology Research Institute in Shanghai,Shanghai 201109,China²

Online:2019-06-14 Published:2019-07-02

摘要/Abstract

摘要： 将深度网络模型应用在图像超分辨率上取得了很大的成功,并且已经证明了在将低分辨率图像重建成高分辨率图像的重建质量上深度网络模型普遍高于传统的算法。为了进一步提高图片的重建质量,文中提出了全局残差递归网络。通过优化经典的残差网络,提出全局残差块特征融合和局部残差块特征融合,让模型产生“自适应”更新权值的思想,改善信息流。结合L1代价函数,ADAM优化器进一步提高了训练的稳定性,并通过DIV2K训练集来训练模型。通过PSNR/SSIM图像重建指标来评价图片重建质量,在SSIM指标中,所提模型最高可达0.94,优于目前最新的深度学习模型(EDSR)的0.92。全局残差递归网络模型有效地提高了图像的重建质量,减少了训练时间,避免了梯度衰减,提高了学习效率。

关键词: ADAM优化器, DIV2K训练集, L1代价函数, 全局残差递归网络, 图像超分辨率

Abstract: The application of the deep network model has achieved great success in image super-resolution,and it has been proven that the reconstruction quality of low-resolution images reconstructed into high-resolution images is gene-rally higher than traditional algorithms.In order to further improve the reconstruction quality of image,a global residual recursive network was proposed .By optimizing the classical residual network,the global residual block feature fusion and the local residual block feature fusion are proposed,which allows the model to generate the idea of adaptive updating weights,and it improves information flow.In combination with the L1 cost function,the ADAM optimizer further improves training stability and trains the model through the DIV2K training set.Through the PSNR/SSIM image reconstruction index,the quality of picture reconstruction is obtained.In the SSIM index,the maximum value is 0.94,which is superior to 0.92 of the current latest deep learning model(EDSR).The global residual recursive network model effectively improves the image reconstruction quality,reduces straining time,effectively avoids gradient attenuation,and improves learning efficiency.

Key words: ADAM optimizer, DIV2K trai-ning set, Global residual recursive network, Image super-resolution, L1 cost function

中图分类号:

TP394

张雷, 胡博文, 张宁, 王茂森. 图像超分辨率全局残差递归网络[J]. 计算机科学, 2019, 46(6A): 230-233. https://doi.org/

ZHANG Lei, HU Bo-wen, ZHANG Ning, WANG Mao-sen. Global Residual Recursive Network for Image Super-resolution[J]. Computer Science, 2019, 46(6A): 230-233. https://doi.org/

参考文献

[1]ALLEBACH J,WONG P W.Edge-directed interpolation[C]∥ICIP.1996.
[2]DONG C,LOY C C,HE K,et al.Learning a deep convolutional network for image super-resolution[C]∥ECCV.2014.
[3]DONG C,LOY C C,TANG X.Accelerating the superresolution convolutional neural network[C]∥ECCV.2016.
[4]HE K,ZHANG X,REN S,et al.Deep residual learning for image recognition[C]∥CVPR.2016.
[5]KINGMA D,BA J.Adam:A method for stochastic optimization[C]∥ICLR.2014.
[6]KIM J,LEE J K,LEE K M.Accurate image super resolution using very deep convolutional networks[C]∥CVPR.2016.
[7]DONG C,LOY C C,TANG X.Accelerating the superresolution convolutional neural network[C]∥ECCV.2016.
[8]GLOROT X,BORDES A,BENGIO Y.Deep sparse rectifier neural networks[C]∥AISTATS.2011.
[9]LEDIG C,THEIS L,HUSZ′A F,et al.Photo-realistic single ima-ge super-resolution using a generative adversarial network[C]∥CVPR.2017.
[10]LIM B,SON S,KIM H,et al.Enhanced deep residual networks for single image super-resolution[C]∥CVPR.2017.
[11]KIM J,LEE J K,LEE K M.Deeply-recursive convolutional network for image super-resolution[C]∥CVPR.2016.
[12]HE K,ZHANG X,REN S,et al.Deep residual learning for image recognition[C]∥CVPR.2016.
[13]SZEGEDY C,IOFFE S,VANHOUCKEET V,et al.Inception-v4,inception-resnet and the impact of residual connections on learning[C]∥AAAI.2017.
[14]KARRAS T,AILA T,LAINE S,et al.Progressive growing of gans for improved quality,stability,and variation[C]∥ICLR.2017.
[15]TIMOFTE R,AGUSTSSON E,GOOL L V,et al.Ntire 2017 challenge on single image super-resolution Methods and results[C]∥CVPRW.2017.
[16]ZHAO H,GALLO O,FROSIO I,et al.Loss functions for neural networks for image processing[J].arXiv:1511.08861,2015.
[17]李键红,吴亚,吕巨建.基于组稀疏表示的在线单帧图像超分辨率算法[J].计算机科学,2018,45(4):312-318.
[18]刘甜甜,包芳勋,张云峰,等.有理分形曲面造型及其在图像超分辨中的应用[J].计算机科学,2018,45(3):35-45.
[19]TAI Y,YANG J,LIU X,et al.Memnet:A persistent memory network for image restoration[C]∥ICCV.2017.

相关文章 6

[1]	周颖, 常明新, 叶红, 张燕. 基于元迁移的太阳能电池板缺陷图像超分辨率重建方法 Super Resolution Reconstruction Method of Solar Panel Defect Images Based on Meta-transfer 计算机科学, 2022, 49(3): 185-191. https://doi.org/10.11896/jsjkx.210100234
[2]	刘遵雄, 朱成佳, 黄稷, 蔡体健. 多跳连接残差注意网络的图像超分辨率重建 Image Super-resolution by Residual Attention Network with Multi-skip Connection 计算机科学, 2021, 48(11): 258-267. https://doi.org/10.11896/jsjkx.201000033
[3]	莫彩网, 常侃, 李恒鑫, 李明鸿, 覃团发. 基于通道间相关性和非局部自相似性的彩色图像超分辨率算法 Color Image Super-resolution Algorithm Based on Inter-channel Correlation and Nonlocal Self-similarity 计算机科学, 2020, 47(6): 138-143. https://doi.org/10.11896/jsjkx.190500047
[4]	李键红,吴亚榕,吕巨建. 基于组稀疏表示的在线单帧图像超分辨率算法 Online Single Image Super-resolution Algorithm Based on Group Sparse Representation 计算机科学, 2018, 45(4): 312-318. https://doi.org/10.11896/j.issn.1002-137X.2018.04.053
[5]	李键红,吕巨建,吴亚榕. 基于最大化自相似性先验的盲单帧图像超分辨率算法 Blind Single Image Super-resolution Using Maximizing Self-similarity Prior 计算机科学, 2018, 45(2): 147-151. https://doi.org/10.11896/j.issn.1002-137X.2018.02.026
[6]	黄华李俊齐春朱世华. 基于通用高斯马尔可夫随机场模型的图像超分辨率重建计算机科学, 2005, 32(11): 195-197.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed