Local version of "White balance" in GIMP to clean up photo of document

Question

I have a photograph of a document with black text on white background.

The photo has a few problems:

• The text is a bit blurry.
• Overall there is noise even in the white areas.
• The background appears not really white but a bit yellow-ish.
• Most important: Some areas of the picture, even the areas that are supposed to be white, are darker than others.

I would like to clean this picture up. I am on Linux.

The "Colors > Auto > White balance" filter in GIMP is producing promising results. However, it does not equalize the background in different areas in the picture.

But if I select only a sub-area of the image, the "White balance" filter actually works better in this area.

So I imagine that gradually applying "White balance" in local areas of the picture would work really great.

I imagine that the "White balance" filter is implemented like this:

1. Collect color statistics of the entire picture.
2. Build a color conversion matrix, and apply it globally.

So what I'd like instead:

1. Collect local color statistics for every area of e.g. 100*100 px.
2. Build local color conversion matrices.
3. Build continuous function of conversion matrices per pixel.
4. Apply locally per pixel.

Do you know anything like this that exists either in GIMP or as an independent piece of software e.g. for the Linux command line?

Answer 1

I assume you have a gradient in overall brightness due to unevben lighting. If that's the case, what you could do is to duplicate the layer with the image and apply the best correction to each part of the image on a different layer: curves for contrast, color balance or desaturate for white balancing.

Then, using layer masks you can paint each region with smooth gradients, in order to blend them together. Desaturation can be used as last step to even a bit the remaining differences.

An alternative approach could be to duplicate the layer (i took a picture with my hand shading half of the card, to reproduce the issue):

Blur it until the text disappears:

And then subtract the resulting layer from the original one select the blending mode as "Grain extract", in order to reverse the unevenness of the lighting:

The lowpass filtering that you achieve by blurring is only effective if changes in lighting are much smoother than the details of text, and it works best with sparser text as it disappears with blurring. Note that in this case, the bold text causes a halo when blurred, which gets inverted when doing the blending. You can try with a larger radius, or try masking it before blurring the duplicate layer.

Also, note that this operation also fixes the white balance as it inverts the color curves.

As a final touch I applied the 'Curves' setting to increase contrast.

It may not be perfect but I'm actually surprised by how that one came out.

Answer 2

You are likely to have some success doing levels in the channels that are darker (if the image looks yellow, lighten the blue channel etc.) You can do this globally. This is likely to be in addition to (before) levels on the *value" channel. In the value channel, you can move the input black and white points closer together, then probably increase the gamma to lighten the paper that's affected by the blur from the print.

While desaturating might help it will leave you with a grey background rather than white if applied before you lift the background levels - it will be useful once the background looks white. Unsharp mask with a fairly low value but large radius can lift areas of white around the text, while a small radius can overcome some of the blur.

I don't recommend threshold unless you absolutely have to provide black-and-white, as greyscale has an anti-aliasing effect.

Once your background is white (rgb 255,255,255) well away from the text, local spot removal might be necessary.

Answer 3

...blending mode as "Grain extract"

"Division" mode also produces nice effects for documents (more contrastful)

Answer 4

Here is an idea without concrete implementation instructions. If someone can flesh this out in a new answer, go ahead!

First you make a copy of the image, e.g. in a new layer - depends on the tool you use.

In this copy, you first do a bit of smoothing with a small radius. This means, every pixel becomes a weighted average of its local surrounding.

Then you apply a median, or a "lower median", with a radius larger than the average character. This means, for every pixel (and every color channel):

• Collect all pixels in the vicinity, e.g. 100 = 10x10 pixels.
• Sort them by brightness.
• Use the 50th brightest color value, as the new color value for the central pixel. Or the 10th darkest color value, for "lower median".

This gives you an estimate of the background. However, it will likely contain "steps" in the colors due to how medians work. Maybe there is something even smarter than these medians.

Optionally, use some smoothing (large radius) to eliminate these steps.

Now subtract this layer from the original image.

Optionally, use "white balance" on the result (if you are working with GIMP).

I don't know which tool I should pick if I wanted to implement all this. So far I used PHP, but I am not convinced that this is the best solution at all.