BeltCrack: the First Sequential-image Industrial Conveyor Belt Crack Detection Dataset and Its Baseline with Triple-domain Feature Learning

News • Datasets Overview • Method • Start • Email

📢 News

• 2025-06-22: The paper is now available on arXiv: http://arxiv.org/abs/2506.17892
• 2025-06-23: BeltCrackDet PyTorch code are released!
• 2026-03-21: Accepted by : https://doi.org/10.1016/j.patcog.2026.113598 🥂🥂🥂
• 2026-03-27：BeltCrack datasets is now officially available on ScienceDB. 🎉🎉🎉

📊 Datasets Overview

Download

• 🔗 Download Link: https://doi.org/10.57760/sciencedb.31181 Unzip Password: cv205

• 📖 Citation: If you use this dataset in your research, please kindly cite our paper:

    @article{HUANG2026113598,
      title = {BeltCrack: the First Sequential-image Industrial Conveyor Belt Crack Detection Dataset and Its Baseline with Triple-domain Feature Learning},
      author = {Jianghong Huang and Luping Ji and Xin Ma and Mao Ye},
      journal = {Pattern Recognition},
      pages = {113598},
      year = {2026},
      issn = {0031-3203},
      doi = {https://doi.org/10.1016/j.patcog.2026.113598},
      url = {https://www.sciencedirect.com/science/article/pii/S0031320326005649}
    }

Caption

• The dataset BeltCrack14ks contains 14,087 images, across 29 sequences. While BeltCrack9kd comprises 9,645 images, from 42 sequences.
• They are captured in real-world industrial environments, including conveyor the belt cracks under multiple perspectives (top-down, bottom-up), varying the lighting conditions from morning strong light to evening low illumination, extreme weather (sunny, rainy, snowy), and dynamic belt moving speeds.

🔑 Method

We propose a tri-path network architecture to implement cross-domain representation learning through Hierarchical Spatial-aware Module (HSM), Aggregative Temporal Module (ATM), and Wavelet-enhanced Frequency-aware Module (WFM). In addition, Residual Compensation Unit (RCU) dynamically mitigates inter-domain representational gaps, while optimizing cross-domain feature fusion.

🔱 Start

Enviroment

You can create your own conda environment for BeltCrackDet based on the following commands:

conda create -n BeltCrackDet python=3.9 
conda activate BeltCrackDet
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
pip install opencv-python==4.11.0.86
pip install einops scikit-learn numpy

Train

You could modify the parameters or paths in the train_BeltCrackDet.py file and run it with the following command (for two GPUs):

CUDA_VISIBLE_DEVICES=0,1 python train_BeltCrackDet.py 

Test

Once training is complete, you could choose best model (usually not 'best_epoch_weights.pth') from "results/beltcrack" to test the performance, and use the following command (for two GPUs):

CUDA_VISIBLE_DEVICES=0,1 python vid_map_coco.py

Visulization

You could choose the mode "predict" in the "predict.py" file to get the results:

python predict.py

The visualization results of our comparative experiments are as follows:

📧 Email

If you have any questions, contact me via email (with the subject of BeltCrack): jianghong@std.uestc.edu.cn

🏷️License

This project is released under the Apache 2.0 license.