scene_text_segmentation

Pixel-wise scene text segmentation (paper) based on DeepLabV3+ paper and its Pytorch implementation.

Results

Qualitative results of English (first four columns) from ICDAR2013 dataset and Korean (fifth to eighth columns) from KAIST dataset. Korean text has been segmented in zero-shot learning, the trained models have never seen the Korean text images.

Installation

Create a conda environmet by installing following packages:

conda install python=3.6 ipython pytorch=0.4 torchvision opencv=3.4.4 tensorboardx mkl=2019 tensorboard tensorflow tqdm scikit-image

• Required packages:
  • Pytorch 0.4
  • OpenCV 3.4.4
  • mkl 2019
  • tqm
  • scikit-image
  • tensorboardX

Train

The path for training dataset should be defined in mypath.py. Then, for instance for ICDAR dataset in dataloaders/datasets the icdar.py refers to that.

• For training ICDAR:
  

bash train_icdar.sh

Test

• For visualizing the heatmaps:
  

visual_hm.py

• For saving the binary text segmentations:
  

test_save_binary.py

• For computing the F1 accuracy:
  

F1_accuaracy_rwi.py

Citation

Please cite this work in your publications if it helps your research:

@article{Rawi19,
       author = {Mohammed Al-Rawi and Dena Bazazian and Ernest Valveny},
       title = {Can Generative Adversarial Networks Teach Themselves Text Segmentation?},
       journal = {IEEE Proceedings of International Conference on Computer Vision Workshops},
       year = {2019}
}