DocClean

Clean your document images with crumpled backgrounds, strains and folds with Deep Neural Networks.

Here is a demo:

You can find the slide deck accompanying this project here.

Installation

For installing docclean is easy. Just run the following:

git clone https://github.com/devanshkv/insight_docclean.git
cd insight_docclean
pip install -r requirements.txt
python3 setup.py install

Documentation

Have a look at our beauitful docs here.

Training the models

The models can be trained using train.py. The usage is as follows:

usage: train.py [-h] -t {cycle_gan,autoencoder} -k KAGGLE_DATA_DIR
                     [-c CLEAN_BOOKS_DIR]
                     [-d DIRTY_BOOKS_DIR] [-e EPOCHS]
                     [-b BATCH_SIZE] [-v]

Quick reference table

Short	Long	Default	Description
-h	--help		show this help message and exit
-t	--type	None	Which model to train
-k	--kaggle_data_dir	None	Kaggle Data Directory
-c	--clean_books_dir	None	Directory containing clean images
-d	--dirty_books_dir	None	Directory containing dirty images
-e	--epochs	100	Number of epochs to train for
-b	--batch_size	16	Batch size
-v	--verbose		Be verbose

Running the inference

Using the trained model the infence can be run using infer.py. The usage is as follows:

usage: infer.py [-h] [-v] [-g GPU_ID] -c DATA_DIR [-b BATCH_SIZE] -t
               {cycle_gan,autoencoder} -w WEIGHTS

Quick reference table

Short	Long	Default	Description
-h	--help		show this help message and exit
-v	--verbose		Be verbose
-g	--gpu_id	0	GPU ID (use -1 for CPU)
-c	--data_dir	None	Directory with candidate pngs.
-b	--batch_size	32	Batch size for training data
-t	--type	None	Which model to train
-w	--weights	None	Model weights

Running the streamlit app

Run,

streamlit run app.py

and the use localhost:8501 to view the app.