|
--- |
|
license: apache-2.0 |
|
tags: |
|
- Scene Text Removal |
|
- Image to Image |
|
library_name: pytorch |
|
--- |
|
|
|
### GaRNet |
|
|
|
This is text-removal model that introduced in the paper below and first released at [this page](https://github.com/naver/garnet). \ |
|
[The Surprisingly Straightforward Scene Text Removal Method With Gated Attention and Region of Interest Generation: A Comprehensive Prominent Model Analysis](https://arxiv.org/abs/2210.07489). \ |
|
Hyeonsu Lee, Chankyu Choi \ |
|
Naver Corp. \ |
|
In ECCV 2022. |
|
|
|
### Model description |
|
|
|
GaRNet is a generator that create non-text image with given image and coresponding text box mask. It consists of convolution encoder and decoder. The encoder consists of residual block with attention module called Gated Attention. |
|
|
|
Gated Attention module has two Spatial attention branch. Each attention branch finds text stroke or its surrounding regions. The module adjusts the weight of these two domains by trainable parameters. |
|
|
|
The model was trained in PatchGAN manner with Region-of-Interest Generation. \ |
|
The discriminator is consists of convolution encoder. Given an image, it determines whether each patch, which indicates text-box regions, is real or fake. |
|
All loss functions treat non-textbox regions as 'don't care'. |
|
|
|
### Intended uses & limitations |
|
|
|
This model can be used for areas that require the process of erasing text from an image, such as concealment private information, text editing.\ |
|
You can use the raw model or pre-trained model.\ |
|
Note that pre-trained model was trained in both Synthetic and SCUT_EnsText dataset. And the SCUT-EnsText dataset can only be used for non-commercial research purposes. |
|
|
|
### How to use |
|
|
|
You can use inference code in [this page](https://github.com/naver/garnet). |
|
|
|
### BibTeX entry and citation info |
|
|
|
``` |
|
@inproceedings{lee2022surprisingly, |
|
title={The Surprisingly Straightforward Scene Text Removal Method with Gated Attention and Region of Interest Generation: A Comprehensive Prominent Model Analysis}, |
|
author={Lee, Hyeonsu and Choi, Chankyu}, |
|
booktitle={European Conference on Computer Vision}, |
|
pages={457--472}, |
|
year={2022}, |
|
organization={Springer} |
|
} |
|
``` |
