Stroke-Based Scene Text Erasing

This repository is a PyTorch implemention of following paper:

Stroke-Based Scene Text Erasing Using Synthetic Data for Training | IEEE Xplore

Zhengmi Tang, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi.

Graduate School of Engineering, Tohoku University.

Requirements

PyTorch==1.8.1
scikit_image==0.18.1
tqdm==4.55.1
torchvision==0.9.1
numpy==1.19.2
opencv-python==4.5.1.48

Training

perpare the training dataset root as:

--train_set
     |i_s
        | 1.jpg
        | 2.jpg
     |mask_t
        | 1.png
        | 2.png
     |t_b
        | 1.jpg 
        | 2.jpg

tune the training parameters in cfg.py. If you want to funetune the model, turn the flag of finetune and resume both in True.
run

python train.py

Testing

perpare the format of testing data as ./examples file shows. The names of txt files should be gt_img_?.txt or res_img_?.txt and the annotation of text bboxes should be quadrilateral.
download our retrained pretrained model. (Note that we retrained our model in a different preprocessing strategy of training data for better visual perception. best.pth for real-world data testing and best_syn.pth for synthetic data testing)
revise the model_path, src_img_dir and src_txt_dir with the right path in test.py
run

python test.py

Evaluation

perpare the names of result images and label images as same name, for example, img_?.png or img_?.jpg
revise the result_path and label_path with the right path in eval.py
run

python eval.py

Citation

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

@article{StrokeErase2021tang,
  author = {Tang, Zhengmi and Miyazaki, Tomo and Sugaya, Yoshihiro and Omachi, Shinichiro},
  journal = {IEEE Transactions on Image Processing},
  title = {Stroke-Based Scene Text Erasing Using Synthetic Data for Training},
  year = {2021},
  volume = {30},
  pages = {9306-9320}
}

Acknowledge

We thank tanimutomo for the excellent code.