DIS-Handwriting-Remover

Handwriting-Removal-DIS

My effort into improving handwriting removal throught the new DIS (Dichotomous Image Segmentation)

Inference

1. Clone the DIS github:

git clone https://github.com/xuebinqin/DIS

1. Download the isnet.pth file from this huggingface model repository and move it into the cloned DIS folder.

2. Replace Inference.py in the cloned DIS folder to the Inference.py of this repository.

3. Change the paths according to your own application (Evaluation data path may be different).

Related Research

AndSonder has also done research and experimentaion on the same subject but using deeplabv3+ to segment the handwriting.

This is a link to his repo: https://github.com/AndSonder/HandWritingEraser-Pytorch

HUGE THANKS to them for providing the segmentation datasets labeled with background blue, printed characters green, and handwriting in red.

Dataset

The original dataset is in Baidu Web Storage and is a segmentation dataset, unlike a background removal dataset.

Therefore, after some processing, I generated a background-removal dataset. It is available in Huggingface: https://huggingface.co/datasets/Inoob/HandwritingSegmentationDataset.

The relavent contents of the repo is listed:

|- train.zip
|- val.zip

After unzipping train.zip and val.zip, the file tree should look like:

|-train
|    |-gt
|    |  |- dehw_train_00714.png
|    |  |- dehw_train_00715.png
|    |  ...
|    |-im
|    |  |- dehw_train_00714.jpg
|    |  |- dehw_train_00715.jpg
|-val
|    |-gt
|    |  |- dehw_train_00000.png
|    |  |- dehw_train_00001.png
|    |  ...
|    |-im
|    |  |- dehw_train_00000.png
|    |  |- dehw_train_00001.png

the gt folder is masks. With the background masked in black, and the handwriting masked as white (a.k.a ground truth data).

the im folder is the normal image of the handwriting dataset.

The code that was used to generate the dataset in the Huggingface Repo is create_masks.py

Training

I used the train_valid_inference_main.py from DIS with my own dataset and training batch size.