Can MaIR’s Locality‑Preserving Mamba Boost Image Restoration?

Paper title: MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration . arXiv PDF: https://arxiv.org/pdf/2412.20066. Source code: https://github.com/XLearning-SCU/2025-CVPR-MaIR

Method

Comparison of visual Mamba scanning strategies.

Overall architecture: three linear stages – shallow feature extraction (3×3 Conv), deep feature extraction (stacked RMG blocks), and reconstruction layer (super‑resolution: 3×3 Conv + pixel‑shuffle; denoising: 3×3 Conv + residual connection). Each RMG contains several RMB+Conv blocks; RMB is a transformer‑style block whose core is VMM, a dual‑branch Mamba block whose core is MaIRM.

MaIRM module: the NSS (Nested S‑shaped Scanning) flattens 2D features into four 1‑D sequences using a four‑direction nested S‑shaped scanning pattern. The SSO (Self‑Similarity Operator) captures long‑term dependencies within each sequence. The SSA (Sequence‑wise Spatial Aggregation) aggregates the four sequences by pooling, channel shuffle, group convolution, unshuffle, chunked attention, and weighted sum to produce the final output.

Loss functions: L1 loss for super‑resolution and Charbonnier loss for denoising.

Experiments

Quantitative evaluation

Super‑resolution: benchmark PSNR/SSIM improvements (exact numbers omitted).

Denoising: quantitative metrics on standard denoising datasets.

Visual quality

Super‑resolution visual comparison shows sharper edges and more faithful texture reconstruction.

Ablation study

Components examined: NSS, SSA, and bandwidth (feature channel width). Removing any component degrades performance, confirming their contribution.

Insights

Improved scanning strategies (NSS) and dedicated aggregation modules (SSA) are effective ways to enhance Mamba‑based image restoration models.

Compared with the predecessor MambaIR, MaIR achieves higher PSNR/SSIM on both super‑resolution and denoising tasks.

Reference

@inproceedings{MaIR,
  title={MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration},
  author={Li, Boyun and Zhao, Haiyu and Wang, Wenxin and Hu, Peng and Gou, Yuanbiao and Peng, Xi},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2025},
  address={Nashville, TN},
  month={jun}
}