Awesome

Selective Scene Text Removal (SSTR)

This is the pytorch implementation of Paper: Selective Scene Text Removal (Accepted at the 34th British Machine Vision Conference(BMVC2023)). The paper is available at this link.

Setup

To set up their environment, please run:
(we recommend to use Anaconda for installation.)

conda env create -n SSTR -f SSTR.yml
conda activate SSTR

Dataset

We used a dataset created by SynthText. Five types of images were used: a background image, a composite image created by SynthText, a text-only image, a text image with a specific word removed, and an ideal output image. The image above shows an example of the specific word "drug" .
In addition, a csv file containing the conditions used by the Selective Word Removal Module is required. For example, if the target words are {dog, cat, car, book}, the condition is 2 if cat is to be removed. The one-hot vector in this case is (0, 1, 0, 0).
For details, check the data structure in the sample_data folder.

Execution scripts

train

(module name)_train.py: Training codes for each module.
selective_scene_text_removal_fine-tuning_train.py: end-to-end fine-tuning code.

test

(Module name)_test.py: Test codes for each module.
selective_scene_text_removal_fine-tuning_test.py: SSTR test code.

Annotation

We refer to SRNet to develop Background Extraction Module. Also, we refer to Conditioned U-Net to develop Selective Word Removal Module. The links for each implementation are as follows.
SRNet : https://github.com/Niwhskal/SRNet
Conditioned U-Net : https://github.com/ws-choi/Conditioned-U-Net-pytorch
Each of these modules may be used under a separate license.

Citation

If you make use of our work, please cite our paper:

@inproceedings{Mitani_2023_BMVC,
author    = {Hayato Mitani and Akisato Kimura and Seiichi Uchida},
title     = {Selective Scene Text Removal},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year      = {2023},
url       = {https://papers.bmvc2023.org/0521.pdf}
}