.png){kind=icon}
.png){kind=icon}
Return
Reference-based Image Super-Resolution with Mamba-Deformable Convolution Networks
.png){kind=icon}
PRE
AI
DOI：10.1007/s11760-025-04985-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reference-based Image Super-Resolution (RefSR) enhances low-resolution (LR) images by transferring texture details from external reference (Ref) images. Existing methods still face challenges in establishing correspondences between LR and Ref images: patch matching and deformable convolution approaches struggle with long-distance correspondences and fail to effectively utilize all potential reference information; Transformer-based methods, while capable of establishing global dependencies between LR and Ref images, suffer from high computational complexity. To address these issues, this paper proposes a novel RefSR method that combines the local precise alignment capability of deformable convolution with the Mamba model's efficient long-distance dependency modeling through an alternating scanning strategy, achieving efficient and accurate reference feature alignment. Additionally, we design a wavelet-based feature modulation module to enhance important high-frequency texture features from Ref images and a reference feature fusion module capable of adaptive filtering and deep integration of reference features. Experiments on multiple benchmark datasets validate the effectiveness of our proposed method both quantitatively and qualitatively.
Keywords:
Reference-based Image Super-Resolution
Texture Transfer
Vision Mamba
Wavelet Transform
Journal
IF:
2.1
Papers:
877
Citations:
4.6K