Improved Reference-Based Super-Resolution with Frequency Information Fusion

As one of the super-resolution (SR) technologies, reference-based super-resolution (RefSR) has greatly improved SR performance by using high-quality reference images as inputs. However, existing RefSR methods have limitations in that global texture and structure information is not sufficiently transferred because they transfer and fuse texture information of reference images based on the similarity between LR-reference images in the spatial domain. To address this problem, this paper proposes methods for transferring and fusing texture information of reference images in the frequency domain. By applying the proposed methods to the existing RefSR method, TTSR, we show that complementary texture information in the spatial and frequency domains can be transferred and utilized together, which can significantly improve SR performance. In particular, the PieAPP value, a perceptual quality indicator of SR images, has improved by more than 10%.