计算机应用 ›› 2021, Vol. 41 ›› Issue (4): 1020-1026.DOI: 10.11772/j.issn.1001-9081.2020081299

所属专题: CCF第35届中国计算机应用大会(CCF NCCA 2020)

• CCF第35届中国计算机应用大会(CCF NCCA 2020) • 上一篇    下一篇

基于拉普拉斯金字塔生成对抗网络的图像超分辨率重建算法

段友祥1, 张含笑1, 孙歧峰1, 孙友凯2   

  1. 1. 中国石油大学(华东)计算机科学与技术学院, 山东 青岛 266580;
    2. 中国石化胜利油田分公司物探研究院, 山东 东营 257000
  • 收稿日期:2020-08-25 修回日期:2020-09-15 出版日期:2021-04-10 发布日期:2020-11-05
  • 通讯作者: 张含笑
  • 作者简介:段友祥(1964—),男,山东东营人,教授,博士,CCF会员,主要研究方向:网络与服务计算、计算机技术在油气领域的应用;张含笑(1995—),女,山东潍坊人,硕士研究生,主要研究方向:人工智能;孙歧峰(1976—),男,山东东营人,讲师,博士,主要研究方向:计算机技术在油气领域的应用;孙友凯(1974—),男,山东陵城人,高级工程师,硕士,主要研究方向:云计算。
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(20CX05017A);中石油重大科技项目(ZD2019-183-006)。

Image super-resolution reconstruction algorithm based on Laplacian pyramid generative adversarial network

DUAN Youxiang1, ZHANG Hanxiao1, SUN Qifeng1, SUN Youkai2   

  1. 1. College of Computer Science and Technology, China University of Petroleum, Qingdao Shandong 266580, China;
    2. Geophysical Research Institute of Sinopec Shengli Oilfield Company Limited, Dongying Shandong 257000, China
  • Received:2020-08-25 Revised:2020-09-15 Online:2021-04-10 Published:2020-11-05
  • Supported by:
    This work is partially supported by the Fundamental Research Funds for the Central Universities (20CX05017A), the Major Scientific and Technological Project of CNPC (ZD2019-183-006).

摘要: 针对目前的图像超分辨率重建算法中存在的大尺度因子的重建效果较差、不同尺度的图像重建均需要单独训练等问题,提出一种基于拉普拉斯金字塔生成对抗网络(GAN)的图像超分辨率重建算法。算法中的生成器使用金字塔结构实现多尺度的图像重建,从而以渐进上采样的方式降低了大尺度因子的学习难度,并在层与层之间使用密集连接加强特征传播,从而有效避免了梯度弥散问题。算法中使用马尔可夫判别器将输入数据映射为结果矩阵,并在训练的过程中引导生成器关注图像的局部特征,从而丰富了重建图像的细节。实验结果表明:在Set5等基准数据集上分别进行放大2倍、4倍、8倍的图像重建时,所提算法的平均峰值信噪比(PSNR)分别达到了33.97 dB、29.15 dB、25.43 dB,平均结构相似性(SSIM)分别达到了0.924、0.840、0.667,相比用于超分辨率重建的卷积神经网络(SRCNN)、深度拉普拉斯金字塔超分辨率重建网络(LapSRN)、用于超分辨率重建的生成对抗式网络(SRGAN)等其他算法有较大提升,且其重建的图像在主观视觉上保留了更多生动的纹理和小颗粒细节。

关键词: 超分辨率重建, 大尺度因子, 密集连接, 拉普拉斯金字塔, 生成对抗网络

Abstract: Concerning the problems of poor reconstructing performance with large-scale factors and requirement of separate training in image reconstruction with different scales in current image super-resolution reconstruction algorithms, an image super-resolution reconstruction algorithm based on Laplacian pyramid Generative Adversarial Network(GAN) was proposed. The pyramid structure generator of the proposed algorithm was used to realize the multi-scale image reconstruction, so as to reduce the difficulty in learning large-scale factors by progressive up-sampling, and dense connection was used between layers to enhance feature propagation, which effectively avoided the vanishing gradient problem. In the algorithm, Markovian discriminator was used to map the input data into the result matrix, and the generator was guided to pay attention to the local features of the image in the process of training, which enriched the details of the reconstructed images. Experimental results show that, when performing 2-times, 4-times and 8-times image reconstruction on Set5 and other benchmark datasets, the average Peak Signal-to-Noise Ratio(PSNR) of the proposed algorithm reaches 33.97 dB, 29.15 dB, 25.43 dB respectively, and the average Structural SIMilarity(SSIM) of the algorithm reaches 0.924, 0.840, 0.667 respectively, outperforming to those of other algorithms such as Super Resolution Convolutional Neural Network(SRCNN), fast and accurate image Super-Resolution with deep Laplacian pyramid Network(LapSRN) and Super-Resolution GAN(SRGAN), and the images reconstructed by the proposed algorithm retain more vivid textures and fine-grained details in subjective vision.

Key words: super-resolution reconstruction, large-scale factor, dense connection, Laplacian pyramid, Generative Adversarial Network (GAN)

中图分类号: