Cross-Scale Dynamic Alignment Network for Reference-Based Super-Resolution.

Image super-resolution aims to recover high-resolution (HR) images from corresponding low-resolution (LR) images, but it is prone to lose significant details in reconstruction progress. Reference-based image super-resolution can produce realistic textures using an external reference (Ref) image, thus reconstructing pleasant images. Despite the remarkable advancement, there are two critical challenges in reference-based image super-resolution. One is that it is difficult to match the correspondence between LR and Ref images when they are significantly different. The other is how the details of the Ref image are accurately transferred to the LR image. In order to solve these issues, we propose improved feature extraction and matching method to find the matching relationship corresponding to the LR and Ref images more accurately, propose cross-scale dynamic correction module to use multiple scale related textures to compensate for more information. Extensive experimental results over multiple datasets demonstrate that our method is better than the baseline model on both quantitative and qualitative evaluations.