SDDuSR: Sparse Feature Matching and Token Dictionary Learning for Dual-Lens Super-Resolution

ICLR 2026 Conference Submission17071 Authors

Keywords: Reference-based Image Super-Resolution, Dual-Lens Super-Resolution, Sparse Feature Matching, Dictionary Learning

Abstract: Dual-Lens Super-Resolution (DuSR) is an application of Reference-based image Super-Resolution (RefSR) in real-world scenarios. Unlike RefSR, DuSR uses the telephoto image as the high-resolution reference image (Ref) and the wide-angle image as the low-resolution image (LR), where LR and Ref share the field of view (FoV) within a certain area. Then, the Ref image is used to assist the LR image in super-resolution. The existing DuSR methods all employ dense feature matching and warping operation to identify and transfer the high-resolution features of the Ref image to the LR image. However, this approach has two key issues: (1) the smooth low-frequency regions in the LR image can achieve good visual effects without any reference, which leads to significant computational redundancy caused by dense feature matching, and (2) due to the inherent limitations of the warping operation, it is not possible to fully utilize the high-resolution features of the Ref image. To address these issues, we propose a DuSR method based on Sparse Feature Matching and Token Dictionary Learning, called SDDuSR. Specifically, we introduce a mask generator to separate the high-frequency regions from the low-frequency regions of the image, and perform feature matching only on the high-frequency regions, which significantly reduces the computational load during the feature matching stage. Moreover, to fully utilize the features of the Ref image, we abstract it into a token dictionary and employ a dictionary learning strategy to assist the LR image in super-resolution. Extensive experiments have demonstrated that our method achieves state-of-the-arts (SOTA) performance in both quantitative and qualitative aspects.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17071