Metadata-Version: 2.1
Name: whitening
Version: 0.2
Summary: Document whitening (foreground separation)
Home-page: https://github.com/rossumai/whitening
Author: Bohumir Zamecnik
Author-email: bohumir.zamecnik@gmail.com
License: MIT
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.7
Classifier: Operating System :: POSIX :: Linux
Description-Content-Type: text/markdown
Provides-Extra: test

# Document whitening (foreground separation)

This package tries to separate text/line foreground and background by 2D median
filter.

<img src="data/IMG_3262.jpg" alt="original" width="250">
<img src="data/IMG_3262_fg.jpg" alt="foreground" width="250">
<img src="data/IMG_3262_bg.jpg" alt="background" width="250">

## Installation

Install from PyPI. Works on Python 3.

```bash
pip install whitening
```

## Example usage

### Python API

It works with images represented as `PIL.Image` or as a numpy array. Images can
be either RGB or grayscale.

```python
import numpy as np
import PIL.Image

from whitening import whiten

# possible to use numpy array as input/output
image = np.asarray(PIL.Image.open('image.jpg'), dtype='uint8')
foreground, background = whiten(image, kernel_size=20, downsample=4)
PIL.Image.fromarray(foreground).save('foreground.jpg', 'jpeg')

# or directly a PIL image
image = PIL.Image.open('image.jpg')
foreground, background = whiten(image, kernel_size=20, downsample=4)
foreground.save('foreground.jpg', 'jpeg')
```

### CLI

It install an entry point called `whiten`.

```bash
# help
$ whiten -h

# whiten an image and save the foreground output
$ whiten input.jpg foreground.jpg

# specify the kernel size
$ whiten input.jpg foreground.jpg -k 100

# work in grayscale instead of RGB (3x faster)
$ whiten input.jpg foreground.jpg -g

# downsample the image 4x (faster, but a bit less precise)
$ whiten input.jpg foreground.jpg -d 4

# save also the background
$ whiten input.jpg foreground.jpg -b background.jpg
```

We assume the original images is a product of foreground and background,
thus we can recover the foreground by dividing the image by the background:
`I = F * B => F = I / B`. We try to approximate the background by 2D median
filtering the original image which suppresses sparse features such as text and
lines.

Select kernel size that's enough for not making artifacts while small enough
to keep computation fast. A good starting point is 50 pixels.

A 9.5 Mpx image can be processed on a MacBook in 15 s, with grayscale and
downsampling 4x the run time can be reduced to 1 s! Quite good results can be
obtained even with kernel size 10 and downsampling 16x.

More info: http://bohumirzamecnik.cz/blog/2015/image-whitening/

## Development

See the `Makefile` for various development tasks.

## License

Author: Bohumír Zámečník <bohumir.zamecnik@gmail.com>

Supported by [Rossum](https://rossum.ai), creating a world without manual data entry.
