Chinese Journal of Liquid Crystals and Displays, Volume. 37, Issue 10, 1334(2022)

Gradient-aware based single image super-resolution

Le ZHOU1, Long XU2, Xiao-yan LIU3, Xin-ze ZHANG2, and Xuan-de ZHANG1、*
Author Affiliations
  • 1School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an 710021,China
  • 2National Astronomical Observatories,Chinese Academy of Sciences,Beijing 100101,China
  • 3College of Science,Xi’an Shiyou University,Xi’an 710065,China
  • show less

    With the application of generative adversarial networks in the field of image super-resolution (SR)?, some perception-driven SR methods can recover SR images with richer texture details, effectively alleviating the over-smoothing problem of the PSNR dominated SR methods. Gradient information is an important representation of image texture. However, few SR methods can make use of this information accurately and efficiently. In this paper, a gradient-aware single image super-resolution(GASR) is proposed, using gradient information better from two aspects. On the one hand, the feature map of gradient domain is used as the convolution kernel imposing on the feature map of image domain, which can effectively avoid the domain conflict caused by the concatenation of feature map of different domains. On the other hand, by elaborating the network details such as convolution kernel size, etc., the image fields output at the corresponding positions of the two branches are consistent with the feel fields of the feature maps in the gradient domain. In addition, the proposed GASR algorithm also effectively reduces the number of parameters and the amount of computation due to the increased demand for network lightweight in practical applications. Compared to SPSR, GASR can achieve the same performance at the cost of about 1/6 of the parameters and 1/10 of the computation of SPSR. On Set14 dataset, LPIPS and PSNR increase by 0.002 2 and 0.217, respectively. The experimental results show that GASR can achieve a good trade-off between texture details and image smoothness. In addition, GASR can not only reconstruct high fidelity SR image, but also alleviate the generation of messy textures.

    Tools

    Get Citation

    Copy Citation Text

    Le ZHOU, Long XU, Xiao-yan LIU, Xin-ze ZHANG, Xuan-de ZHANG. Gradient-aware based single image super-resolution[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(10): 1334

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Articles

    Received: Mar. 11, 2022

    Accepted: --

    Published Online: Oct. 10, 2022

    The Author Email: Xuan-de ZHANG (zhangxuande@sust.edu.cn)

    DOI:10.37188/CJLCD.2022-0083

    Topics