SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures

Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Shengyong Chen

Abstract

Structural health monitoring relies heavily on the accurate detection of cracks in infrastructure such as roads, bridges, and pipelines. However, existing computer vision methods often struggle to balance computational efficiency with the ability to capture the complex, irregular geometry of cracks against noisy backgrounds. This paper introduces SCSegamba, a lightweight architecture based on State Space Models (SSMs). By incorporating a novel Structure-Aware Scanning Strategy (SASS) and Gated Bottleneck Convolutions (GBC), SCSegamba achieves state-of-the-art performance with only 2.8 million parameters. Experimental results across multiple benchmarks demonstrate that the model provides superior continuity in crack segmentation and is capable of real-time deployment on resource-constrained edge devices.

Introduction

Detecting cracks in structural surfaces is a fundamental yet challenging task. Cracks are characterized by thin, winding topologies and are often obscured by environmental noise, such as shadows, stains, or complex surface textures. Historically, two primary deep learning approaches have dominated this field:

  1. Convolutional Neural Networks (CNNs): While effective at extracting local features, CNNs suffer from a limited receptive field. This often leads to fragmented segmentation results where long, continuous cracks are perceived as disconnected segments.
  2. Vision Transformers (ViTs): Transformers excel at capturing global context and long-range dependencies. However, their self-attention mechanism scales quadratically with image resolution, making them computationally prohibitive for real-time use on inspection drones or mobile robots.

SCSegamba bridges this gap by utilizing the Vision Mamba (SSM) framework, which offers the global modeling capabilities of Transformers with the linear computational complexity of CNNs.

Methodology

The SCSegamba architecture is designed specifically to address the unique morphological characteristics of structural cracks through three primary innovations:

1. Structure-Aware Scanning Strategy (SASS)

Standard Vision Mamba models typically process images using simple horizontal or vertical scanning paths. Because cracks are inherently multi-directional and often diagonal, these standard paths can break the semantic continuity of the crack. SASS introduces "diagonal snake" scanning paths, allowing the model to "follow" the path of a crack regardless of its orientation. This ensures that the network maintains a consistent understanding of the crack's structure across the entire image.

2. Gated Bottleneck Convolution (GBC)

To maintain a lightweight profile, the authors employ low-rank approximations to map high-dimensional data into lower dimensions. The GBC module enhances this by adding a gating mechanism. This mechanism acts as a learned filter, dynamically suppressing background noise (such as tire marks or lighting variations) while highlighting the specific morphology of the crack. This is critical for reducing false positives in real-world inspection scenarios.

3. Multi-scale Feature Segmentation (MFS) Head

Rather than utilizing a heavy, complex decoder, SCSegamba employs a streamlined Multi-layer Perceptron (MLP)-based head. This component aggregates features from multiple layers of the network to produce a high-resolution segmentation map without the significant computational overhead associated with traditional decoders.

Experimental Results

Datasets and Benchmarks

The model was rigorously evaluated on four diverse crack benchmarks:

  • Crack500: Bitumen/asphalt cracks (3,368 images).
  • DeepCrack: Multi-surface cracks including cement and brick (537 images).
  • CrackMap: Complex road cracks in high resolution (120 images).
  • TUT: A highly challenging dataset featuring eight different scenarios, including turbine blades and metal pipes, characterized by heavy background interference (1,408 images).

Performance Metrics

SCSegamba demonstrated exceptional performance, particularly on the challenging TUT dataset, where it achieved an F1 score of 0.8390 and a mean Intersection over Union (mIoU) of 0.8479. These figures outperform nine other state-of-the-art (SOTA) methods, including much larger models.

Efficiency and Deployment

The primary advantage of SCSegamba is its efficiency. With only 2.8 million parameters and a computational load of 18.16G FLOPs, it is significantly smaller than competitors like DTrCNet (63.4M parameters). In practical deployment tests using a Turtlebot4 Lite, the model achieved an inference speed of approximately 32 FPS (0.0313 seconds per frame), proving its readiness for real-time autonomous inspection.

Discussion and Conclusion

SCSegamba proves that high-accuracy structural segmentation does not require massive parameter counts. By optimizing the scanning topology of the Mamba architecture to match the physical structure of the target (cracks), the researchers achieved a "structure-aware" model that excels at maintaining connectivity in thin features.

While the current results are promising, future research will focus on integrating multi-modal data—such as infrared or depth sensing—to improve performance in low-visibility environments. Additionally, further optimization of the Visual State Space (VSS) design may eventually allow the model to run entirely on a robot's onboard processor, eliminating the need for external server communication.

Reference: Liu, H., Jia, C., et al. (2025). SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures. arXiv:2503.01113v3.

Summary updated

BibTeX
@misc{liu2025scsegamba,
  author = {Hui Liu and Chen Jia and Fan Shi and Xu Cheng and Shengyong Chen},
  title = {{SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures}},
  year = {2025},
  eprint = {2503.01113},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.01113},
}
Figure