Structure Preserving Image Super-Resolution Towards Big Data

With the rapid development of big data, single image super-resolution (SISR) has the disadvantage of not being able to integrate with big data due to the inability to fully utilize massive information. Therefore, reference-based image super-resolution (RefSR) technology has emerged. In this paper, we propose a structure-preserving reference-based image super-resolution (SPRefSR) method to address the problem of the inability to reuse the structural information of reference images and effectively restore the self-structure of low resolution image in existing reference-based image super-resolution methods. Our method effectively improves the performance of RefSR by designing the wavelet downsampling block, self-structure branch, and structure losses. The experimental results on CUFED5, Sunhays80, Urban100 and Manga109 benchmarks demonstrate that our proposed framework outperforms related RefSR methods in terms of LPIPS, PSNR, and SSIM metrics.