Abstract
Single image super-resolution (SISR) aims to recover a high-resolution image from a low-resolution input. Despite recent advancements, existing methods often fail to fully exploit self-similarities across image scales. In this paper, we introduce the Deformable Non-Local (D-NL) attention module, integrated into a recurrent neural network. The D-NL attention mechanism leverages deformable convolutions to better capture pixel-wise correlations and long-range self-similarities. Additionally, we propose a Multi-Scale Channel Attention Module (MS-CAM) and a Multi-Branch Weighted Feature Fusion (MWFF) cell to enhance feature fusion, effectively identifying and combining features with distinct semantics and scales. Experimental results on benchmark datasets demonstrate that our approach, DNLN, significantly outperforms state-of-the-art methods, underscoring the effectiveness of exploiting long-range self-similarities for SISR.
| Original language | English |
|---|---|
| Article number | 105721 |
| Pages (from-to) | 1-8 |
| Number of pages | 8 |
| Journal | Image and Vision Computing |
| Volume | 162 |
| DOIs | |
| Publication status | Published - Oct 2025 |
Keywords
- Deformable non-local attention
- Multi-scale channel attention module
- Multi-branch weighted feature fusion
Fingerprint
Dive into the research topics of 'DNLN: image super-resolution with deformable non-local attention and multi-branch weighted feature fusion'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver