Research map / Image generation / Super-resolution
TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TSP-Mamba, which finds a locally optimal (TSP-style) scan path per image patch before a global Mamba scan, improving lightweight single-image super-resolution over MambaIR.
Problem
Fixed 2D-to-1D scanning in Mamba breaks local 2D neighborhoods, so local structure is aggregated poorly; stacking more Mamba layers raises cost, hurting lightweight and real-time image super-resolution.
Approach
Treats each local image patch as a TSP instance over pixels with cosine-distance edge costs, selects the minimum-cost traversal path, and aggregates along it with a bidirectional LSTM (TSP-LSTM) with distance embeddings before the global 2D-SSM scan.
Key results
- TSP-Mamba surpasses MambaIR by up to 0.71 dB on Manga109 at x2 in lightweight SISR, with similar parameters and FLOPs.
- On classic SISR (DF2K), TSP-Mamba-L achieves 0.32/0.37/0.39 dB gains over MambaIR-L at x2/x3/x4 scales.
- On unseen depth-map SR (NYU-V2), TSP-Mamba improves over MambaIR by 0.96 dB at x4 without depth-specific training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TSP-based adaptive local propagation | MambaIR, lightweight SISR · Urban100 ×4 | PSNR: 26.75 -> 27.29 | Helped |
| add TSP-based adaptive local propagation | MambaIR, lightweight SISR · Manga109 ×4 | PSNR: 31.26 -> 31.95 | Helped |
| use negative cosine distance rather than L1 or L2 for scan-path cost (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.55 -> 26.61; 26.54 -> 26.61 | Helped |
| use adaptive TSP scan rather than uniform vertical or horizontal scan (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.31 -> 26.61; 26.38 -> 26.61 | Helped |
| increase local scan patch size from 2 to 4 (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.37 -> 26.61 | Helped |
| increase local scan patch size from 3 to 4 (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.43 -> 26.61 | Helped |
| increase local scan patch size from 4 to 5 (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.61 -> 26.61 | No clear effect |
| add distance embedding and bidirectional propagation (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.16 -> 26.61 | Helped |
| add bidirectional propagation (ablation) | TSP-Mamba-Small with distance embedding · Urban100 ×4 | PSNR: 26.38 -> 26.61 | Helped |
| add distance embedding (ablation) | TSP-Mamba-Small with bidirectional propagation · Urban100 ×4 | PSNR: 26.25 -> 26.61 | Helped |
| use TSP-LSTM rather than Conv1D for local aggregation (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.49 -> 26.61 | Helped |
| use TSP-LSTM rather than RNN for local aggregation (ablation) | TSP-Mamba-Small · Urban100 ×4 | PSNR: 26.44 -> 26.61 | Helped |
Surprising findings
- Adds only about 1-2 GFLOPs for 720P outputs while improving over MambaIR, despite a path search step per patch.
Limitations
- Distance measurements may be unreliable for degraded or noisy images, affecting path search.
- Simplified TSP (4-neighbor moves, fixed start) restricts the search space and is not a full TSP solution.
- Exhaustive path enumeration grows rapidly with patch size, limiting the patch size used.
Experimental setup
Datasets: DIV2K DF2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 NYU-V2
Metrics: PSNR SSIM RMSE FLOPs parameters inference time memory
Baselines: LAPAR-A LatticeNet SwinIR-light DLGSA-light Omni-SR SRFormer HiT-SR ATD-S MambaIR EDSR FALSR-A IMDN
Builds on: MambaIR Mamba fast global image smoothing
Closest papers by meaning
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.87)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.87)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.86)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.86)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.85)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.