Research on single image super-resolution algorithm based on region perception

In recent years, adversarial network-based research has made great progress in single-image super-resolution, In particular, the Enhanced Super-Resolution Generative adversarial networks (ESRGAN) can generate realistic and natural high-resolution images from low-resolution images. However, in the region with rich texture details, the super-resolution results generated by the existing algorithms are often unreal or inaccurate, and often produce artifacts. In order to improve the perception quality of detail texture region, residual paths are added to ESRGAN network structure to enhance feature extraction. To enhance the discriminator's discrimination ability for texture details, we propose a super-resolution training strategy based on region perception. To reduce artifacts, we analyze the source of artifacts and construct artifact loss according to the local statistical information of artifacts. Experiments show that the proposed super-resolution network based on regional awareness achieves better perception quality than the GAN-based SISR presented in recent years.