Research map / Detection and segmentation / CNN features
SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures
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 SCSegamba, a lightweight vision Mamba for pixel-level crack segmentation, using a Structure-Aware Visual State Space block with gated bottleneck convolution and a diagonal-snake scanning strategy (SASS), reaching 2.8M parameters and SOTA F1/mIoU on four crack datasets.
Problem
Pixel-level structural crack segmentation across diverse materials and cluttered backgrounds, where CNNs and Transformers are either limited in receptive field or too computationally heavy, and existing Mamba models do not model crack morphology and texture well.
Approach
A Structure-Aware VSS block combining a lightweight gated bottleneck convolution (GBC) for crack morphology with a four-route structure-aware scan (two parallel snake plus two diagonal snake paths) for topology and texture continuity, plus a multi-scale MLP-based segmentation head (MFS).
Key results
- On the multi-scenario TUT dataset, SCSegamba reaches F1 0.8390 and mIoU 0.8479, beating the next best method by 2.21% F1 and 1.74% mIoU.
- With 2.80M parameters, 18.16G FLOPs and 37MB size at 512 input, it is 52.54% smaller in parameters than the next lightest compared method.
- On DeepCrack it surpasses the next best method by 1.50% F1 and 1.09% mIoU; on CrackMap by 2.06% F1 and 1.65% mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCSegamba instead of MambaIR | Crack segmentation · Crack500 test | F1: 0.7435 -> 0.7553 | Helped |
| use SCSegamba instead of MambaIR | Crack segmentation · DeepCrack test | F1: 0.8975 -> 0.9110 | Helped |
| use SCSegamba instead of SimCrack | Crack segmentation · CrackMap test | F1: 0.7523 -> 0.7678 | Helped |
| use SCSegamba instead of SimCrack | Crack segmentation · TUT test | F1: 0.8208 -> 0.8390 | Helped |
| replace the UNet head with MFS (ablation) | SCSegamba · TUT | F1: 0.8260 -> 0.8390 | Helped |
| replace the Ham head with MFS (ablation) | SCSegamba · TUT | F1: 0.7909 -> 0.8390 | Helped |
| replace the SegFormer head with MFS (ablation) | SCSegamba · TUT | F1: 0.8172 -> 0.8390 | Helped |
| use SASS instead of parallel scanning (ablation) | SCSegamba · TUT | F1: 0.8330 -> 0.8390 | Helped |
| use SASS instead of diagonal scanning (ablation) | SCSegamba · TUT | F1: 0.8320 -> 0.8390 | Helped |
| use SASS instead of parallel snake scanning (ablation) | SCSegamba · TUT | F1: 0.8291 -> 0.8390 | Helped |
| use SASS instead of diagonal snake scanning (ablation) | SCSegamba · TUT | F1: 0.8365 -> 0.8390 | Helped |
| add GBC to SAVSS (ablation) | SCSegamba · TUT | F1: +1.57% | Helped |
Surprising findings
- Adding the residual connection in SAVSS improved mIoU by 2.47% while changing F1 by only 0.13%.
- The MLP-based segmentation head (MFS) outperformed UNet, Ham and SegFormer heads while being comparable in size.
Limitations
- Future work is needed to incorporate multi-modal cues to improve segmentation quality.
- Performance on Crack500 and DeepCrack is still modest relative to some Transformer baselines in some metrics.
- Real-world edge deployment is only referred to the appendix and not discussed in main text.
Experimental setup
Datasets: Crack500 DeepCrack CrackMap TUT
Benchmarks: Crack500 DeepCrack CrackMap TUT
Metrics: ODS OIS precision recall F1 mIoU FLOPs Params Model Size
Baselines: RIND SFIAN CTCrackSeg DTrCNet Crackmer SimCrack CSMamba PlainMamba MambaIR
Builds on: Mamba VMamba Vision Mamba PlainMamba DySample
Closest papers by meaning
- Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models (NeurIPS 2026 · similarity 0.85)
- Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation (CVPR 2024 · similarity 0.85)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.84)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.82)
- TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond (CVPR 2025 · similarity 0.82)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.82)
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.