This is a minimal implementation of Scene text removal via cascaded text stroke detection and erasing. This github repository is for studying on image in-painting for scene text erasing. Thank you :)
Python 3.7 or later with all requirements.txt dependencies installed, including torch>=1.6. To install run:
$ pip install -r requirements.txt
This model has u-net sub modules.
Gd detects text stroke image Ms with I and M. G'd detects more precise text stroke M's.
Similarly, Gr generates text erased image Ite, and G'r generates more precise output I'te.
Not to be confused, I renamed the names.
I : Input Image (with text)
Mm : Text area mask (M in the model)
Ms : Text stroke mask; output of Gd
Ms_ : Text stroke mask; output of G'd
Msgt : Text stroke mask ; ground truth
Ite : Text erased image; output of Gr
Ite_ : Text erased image; output of G'r
Itegt: Text erased image; ground truth
You need to prepare background images in backs directory and text binary images in font_mask directory.
[part of background image sample, text binary image sample]
Executing python create_dataset.py will automatically generate I, Itegt, Mm, Msgt data.
(If you already have I, Itegt, Mm, Msgt, you can skip this section)
├─dataset
│ ├─backs
│ │ # background images
│ └─font_mask
│ │ # text binary images
│ └─train
│ │ └─I
│ │ └─Itegt
│ │ └─Mm
│ │ └─Msgt
│ └─val
│ └─I
│ └─Itegt
│ └─Mm
│ └─Msgt
I generated my dataset with 709 background images and 2410 font mask. I used 17040 pairs for training and 4260 pairs for validation.
Thanks for helping me gathering background images [sina-Kim](sina-Kim (github.com)).
All you need to do is:
python train.pyFrom the left
I, Itegt, Ite, Ite_, Msgt, Ms, Ms_
- Epoch 2
- Epoch 5
- Epoch 10
- Epoch 30
- Epoch 50
- Epoch 120
These are not good enough for real task. I think the reason is lack of dataset and simplicity. But, it was a good experience for me to implement the paper.
If you are having a trouble to run this code, please use issue tab. Thank you.