MambaU-Light: a lightweight model for bridge surface crack segmentation based on joint distillation of multiple models
Source abstract
Abstract
MambaU-Light is a lightweight crack segmentation model aimed at boosting the efficiency and accuracy of bridge surface crack detection, especially under complex settings and resource-constrained scenarios. The technique employs UNet++, TransUNet, and Mamba-UNet as teacher models and uses multi-model joint distillation to transmit their knowledge to the student model, MambaU-Light. Structurally based on UNet, it includes DSCA (Depthwise Separable Channel Attention) and VSS (Visual State Space) modules to reduce parameters and computational complexity while retaining segmentation accuracy dramatically. Experimental results demonstrate that MambaU-Light delivers higher performance with Precision at 82.21%, Recall at 83.03%, F1-score at 81.89%, and mIoU at 84.32%, with a parameter size of only 3.66 M. The study underlines its effectiveness and promise for real-world fracture identification, delivering a novel bridge maintenance solution.
Source: openalex_inverted_index · Quality: baseline
Research context
OpenAlex topics: Infrastructure Maintenance and Monitoring · Concrete Corrosion and Durability · Asphalt Pavement Performance Evaluation
SciVX discovery tags are terminology-based. OpenAlex topics are source-provided classifications.
Affiliations
Xi’an Jiaotong University
Shaanxi College of Communications Technology
Yangzhou University
The University of Melbourne
Central South University
Funding
National Natural Science Foundation of China · Key R&D Plan of Shaanxi Province
References (50)
Show source-provided references
1. 10.1061/(ASCE)ST.1943-541X.000253510.1061/(ASCE)ST.1943-541X.0002535 ↗
2. 10.3390/s2010277810.3390/s20102778 ↗
3. 10.1109/ACCESS.2025.353283210.1109/ACCESS.2025.3532832 ↗
4. 10.1111/mice.1256410.1111/mice.12564 ↗
5. 10.1016/j.aej.2017.01.02010.1016/j.aej.2017.01.020 ↗
6. 10.1155/2018/629049810.1155/2018/6290498 ↗
7. 10.1016/j.cageo.2013.04.00810.1016/j.cageo.2013.04.008 ↗
8. 10.1016/j.aei.2015.01.00810.1016/j.aei.2015.01.008 ↗
9. 10.1145/306538610.1145/3065386 ↗
10. 10.1111/mice.1279310.1111/mice.12793 ↗
11. 10.1111/mice.1226310.1111/mice.12263 ↗
12. 10.1061/(ASCE)CP.1943-5487.000091810.1061/(ASCE)CP.1943-5487.0000918 ↗
13. 10.1007/978-3-319-24574-4_2810.1007/978-3-319-24574-4_28 ↗
14. 10.3390/app1201039510.3390/app12010395 ↗
15. 10.1109/ICASSP40776.2020.905340510.1109/ICASSP40776.2020.9053405 ↗
16. 10.1016/j.conbuildmat.2021.12626510.1016/j.conbuildmat.2021.126265 ↗
17. Dosovitskiy A Beyer L Kolesnikov A et al. An image is worth 16x16 words: transformers for image recognition at scale. 2020.
18. 10.1109/ICCV48922.2021.0006010.1109/ICCV48922.2021.00060 ↗
19. 10.1016/j.autcon.2022.10431610.1016/j.autcon.2022.104316 ↗
20. Zhu X, Su W, Lu L, et al. Deformable DETR: deformable transformers for end-to-end object detection. arXiv preprint arXiv. 2020.
21. 10.1016/j.autcon.2023.10492910.1016/j.autcon.2023.104929 ↗
22. 10.1016/j.jobe.2023.10788910.1016/j.jobe.2023.107889 ↗
23. 10.1016/j.autcon.2022.10427510.1016/j.autcon.2022.104275 ↗
24. 10.1109/TCOMM.2021.312496110.1109/TCOMM.2021.3124961 ↗
25. 10.1177/1475921722113973010.1177/14759217221139730 ↗
26. 10.1016/j.jobe.2023.10720010.1016/j.jobe.2023.107200 ↗
27. 10.1111/mice.1310310.1111/mice.13103 ↗
28. 10.1016/j.autcon.2024.10558510.1016/j.autcon.2024.105585 ↗
29. Zhu M, Gupta S. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878. 2017.
30. 10.1016/j.autcon.2022.10460010.1016/j.autcon.2022.104600 ↗
31. 10.1007/s11263-021-01453-z10.1007/s11263-021-01453-z ↗
32. Zhou Z, Siddiquee MMR, Tajbakhsh N, et al. A nested u-net architecture for medical image segmentation. Deep learning in medical image analysis and multimodal learning for clinical decision support: 4th international workshop, DLMIA 2018. 8th international workshop, ML-CDS 2018; September 20, 2018; Granada, Spain. Springer International Publishing; 2018. Vol. 11045. p. 3–11.
33. Chen J. TransUNet: transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306. 2021.
34. Wang Z, Zheng J-Q, Zhang Y, et al. Mamba-UNet: UNet-Like Pure Visual Mamba for Medical Image Segmentation. arXiv preprint arXiv:2402.05079. 2024.
35. Zhu L, Liao B, Zhang Q, et al. Vision mamba: efficient visual representation learning with bidirectional state space model. arXiv:2401.09417. 2024.
36. 10.1109/TITS.2016.255224810.1109/TITS.2016.2552248 ↗
37. Özgenel ÇF. Concrete crack segmentation dataset. Mendeley Data. 2019;1.
38. 10.1109/TIP.2018.287896610.1109/TIP.2018.2878966 ↗
39. 10.1186/s40537-019-0197-010.1186/s40537-019-0197-0 ↗
40. Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning; Lille, France. p. 448–456.
41. Cortes C, Mohri M, Rostamizadeh A. L2 regularization for learning kernels. arXiv preprint arXiv:1205.2653. 2009.
42. Loshchilov I, Hutter F. Decoupled weight decay regularization. In: International Conference on Learning Representations; April 24–26, 2017; Toulon, France. 2017.
43. Wang A, Chen H, Lin Z, et al. Repvit: revisiting mobile cnn from vit perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Seattle, USA; 2024. p. 15909–15920.
44. 10.1007/978-3-319-67558-9_2810.1007/978-3-319-67558-9_28 ↗
45. 10.1109/TITS.2023.324808910.1109/TITS.2023.3248089 ↗
46. Tan M, Le QV. EfficientNetV2: smaller models and faster training. arXiv preprint arXiv:2104.00298. 2021.
47. 10.1007/978-3-030-01261-8_2010.1007/978-3-030-01261-8_20 ↗
48. 10.1109/ICCV.2019.0014010.1109/ICCV.2019.00140 ↗
49. 10.1080/15732479.2019.168070910.1080/15732479.2019.1680709 ↗
50. 10.1016/j.engfailanal.2023.10723710.1016/j.engfailanal.2023.107237 ↗