LECSFormer

Refined Crack Detection Via LECSFormer for Autonomous Road Inspection Vehicles

Abstract:

Due to the rising cost of human resources in road maintenance and the pursuit of efficiency, autonomous road inspection vehicles are developed for intelligent detection of road disease to prevent severe traffic disasters in the early stages. Nevertheless, as a prevalent road disease, road cracks are diverse and susceptible to shadows, weather changes, and noise in data acquisition. Moreover, they usually appear with thin shapes that are hard to detect correctly by existing methods. To handle this problem, more details of the road cracks need to be better analyzed. In this article, we propose a refined road crack detection method named locally enhanced cross-shaped windows transformer (LECSFormer), which adopts a delicate design of the encoder-decoder structure. The encoder employs window-based transformer blocks to model long-range dependencies. Each transformer block ensembles the locally enhanced module to enrich the local contextual information, and token shuffle operation is applied to build cross-windows connections. The decoder uses dense connections to fuse multi-scale information, and the feature fusion module fuses hierarchical features and reweights them by the channel attention mechanism. The proposed method outperforms other state-of-the-art methods with ODS of 0.963, 0.917, 0.952, and 0.953 on four challenging datasets, CrackTree260, CrackLS315, Stone331, and CRKWH100. It can accurately detect cracks in road surfaces and support intelligent preventive maintenance of roads.

Crack detection for autonomous road inspection vehicles

The proposed Locally Enhanced Cross-Shaped Windows Transformer Architecture consisted of the encoder, decoder, and feature fusion model (FFM),where the LECSFormer blocks are composed of a LECSFormer block and a Shuffle LECSFormer block.

Results

CrackTree260

Model	ODS	OIS	AP	#param.
Ours	0.963	0.969	0.969	16.34M

CrackLS315

Model	ODS	OIS	AP	#param.
Ours	0.917	0.927	0.915	16.34M

Stone331

Model	ODS	OIS	AP	#param.
Ours	0.952	0.970	0.960	16.34M

Get started

Training

python train.py --root_path path/datasets/ --cfg configs/config.yaml

Testing

python test.py --root_path path/datasets/ --cfg configs/config.yaml --output_dir output/

Citation

If our code or models help your work, please cite our papers:

@ARTICLE{9878251,
  author={Chen, Junzhou and Zhao, Nan and Zhang, Ronghui and Chen, Long and Huang, Kai and Qiu, Zhijun},
  journal={IEEE Transactions on Intelligent Vehicles}, 
  title={Refined Crack Detection Via LECSFormer for Autonomous Road Inspection Vehicles}, 
  year={2022},
  volume={},
  number={},
  pages={1-13},
  doi={10.1109/TIV.2022.3204583}}

Acknowledgement

Swin-Unet

CSWin-Transformer

Shuffle Transformer

Thanks for the great implementations!

ZhaoNan1/LECSFormer