How MaIR Advances Image Restoration with a Locality‑Preserving Mamba Architecture

Paper title: MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration

Paper link: https://arxiv.org/pdf/2412.20066

Source code: https://github.com/XLearning-SCU/2025-CVPR-MaIR

Method

Scanning strategies: four visual Mamba scanning patterns are compared (horizontal, vertical, diagonal, and reverse‑diagonal).

Overall architecture: three linear stages – shallow feature extraction (3×3 convolution), deep feature extraction (multiple RMG blocks), and reconstruction. For super‑resolution the reconstruction layer uses a 3×3 convolution followed by pixel‑shuffle; for denoising it uses a 3×3 convolution with a residual connection. Each RMG block contains several RMB+Conv modules; RMB is a transformer‑style block whose core is a VMM (dual‑branch Mamba). The VMM embeds the proposed MaIRM module.

MaIRM module: (a) NSS (Nested S‑shaped) flattens 2‑D features into four 1‑D sequences and scans them in four directions using a nested S‑shaped order; (b) SSO (Self‑Similarity Operator) captures long‑range dependencies across the sequences; (c) SSA (Sequence‑wise Spatial Aggregation) aggregates the four sequences by pooling, channel shuffle, group‑convolution unshuffle, chunked attention weighting, and weighted summation to produce the final output.

Loss functions: L1 loss for super‑resolution tasks; Charbonnier loss (a smooth L1 variant) for denoising tasks.

Experiments

Quantitative evaluation

Super‑resolution: reported PSNR/SSIM improvements on standard benchmarks; visual comparison images illustrate sharper textures.

Denoising: quantitative gains across multiple noise levels; visual results show reduced artifacts.

Quality assessment

Super‑resolution visual quality: subjective evaluation demonstrates clearer edges and better texture fidelity.

Ablation study

NSS: removing the nested S‑shaped scanning reduces performance, confirming its role in preserving locality.

SSA: disabling sequence‑wise spatial aggregation harms the model’s ability to fuse long‑range information.

Bandwidth: varying the channel bandwidth of the VMM shows a trade‑off between accuracy and computational cost.

Summary and Insights

Refining scanning strategies and introducing dedicated modules (NSS, SSO, SSA) in Mamba‑based image restoration models effectively preserves locality and continuity, leading to higher restoration quality.

Compared with the earlier MambaIR model, MaIR achieves superior PSNR/SSIM on both super‑resolution and denoising benchmarks, as illustrated in the comparative figure.

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}
}