Home

Awesome

PERT

The offical implementation of PERT: What is the Real Need for Scene Text Removal? Exploring the Background Integrity and Erasure Exhaustivity Properties (TIP2023). PERT progressively erases the text region with a balanced multi-stage erasure and Region-base Modification Strategy. The model is simple to be implemented (light weight) and can be easily developed.

News

10/8 The model is released
10/8 The code is released.

Performance

The Detailed Performance on SCUT-EnsText

DatasetModelPSNRMSSIMMSEAGEpEPspCEPs
SCUT-EnsTextPaper33.6297.000.00132.18500.01350.0088
SCUT-EnsTextThis implementation34.1297.060.00122.12990.01250.0080

The BI and EE property on SCUT-EnsText

DatasetModelBIEE
SCUT-EnsTextPaper63.5580.00
SCUT-EnsTextThis implementation64.7379.46

Preparetion

Requirement

python=3.6.9 & torch==1.8.1+cu111
It is the best to use our provided Dockerfile for quick start.

Dataset

Download SCUT-EnsText and put it into your folder.

Trained model download

Download our trained model from Google and put it into your folder.

Evaluation

bash test_Pert.sh

Evaluation BI-Metric

python evaluatuion_bi.py --target_path /path_to_output --gt_path /path_to_label --BI True

Evaluation EE-Metric

python evaluatuion_bi.py --target_path /path_to_output --gt_path /path_to_label --BI False

Train

bash train_Pert.sh

Citation

If you find our method useful for your reserach, please cite

@ARTICLE{10214243,
  author={Wang, Yuxin and Xie, Hongtao and Wang, Zixiao and Qu, Yadong and Zhang, Yongdong},
  journal={IEEE Transactions on Image Processing}, 
  title={What is the Real Need for Scene Text Removal? Exploring the Background Integrity and Erasure Exhaustivity Properties}, 
  year={2023},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TIP.2023.3290517}}