How to remove black borders from scanned images?

[izavorin]

I have a bunch of black and white TIFFs which are scans of old documents. Some of them have black borders of various sizes and locations: some around all 4 sides, some only on 1 or 2 sides, some borders thick, some thin. Is it possible to use ImageMagick to remove those?

I tried "-trim" on a couple of them but that produced no effect. Maybe was not using it right. Don't think "-fuzz" will help since images are black and white. What else can I do?

Thx

[GreenKoopa]

Assuming the level of noise isn't too high, IM could do this without too much effort. Grayscale or true 1-bit B&W? If you can post a sample document, even just a mostly empty page to give us an idea of your scanner, someone would probably write the initial command for you.

[izavorin]

The content of the images is a bit sensitive, so can I just post samples with the original borders but with the text area whited out?

[GreenKoopa]

Sure, as long as the nature of the images doesn't change. Things such as 1-bit vs grayscale vs color, resolution (DPI), file format, etc. Are your scans all similar in these regards? Interesting for a crop is both sides of the border, so leave the page margins as much as you can. Do these need to be rotated to straighten? That is a much more difficult task to automate.

[GreenKoopa]

-trim is the right track, and you are right that -fuzz isn't enough to overcome the noise of many scans. There are other ideas already documented here:
http://www.imagemagick.org/Usage/crop/#trim_blur

[fmw42]

I think we will need a (sanitized) example to see what the issue might be and how best to recommend you proceed.

[snibgo]

It's often useful to add a black border before doing the trim. It may be necessary to do some heavy processing to find the black edge. In that case, a trick is to create a mask, so the heavy processing has no impact on the useful areas of the scan.

There have been similar examples on this forum.

[izavorin]

I am having trouble uploading the samples to my online repository so that I can link to them in the post. But in the meantime... Is there a way in IM to compute (black) connected components for a bitonal image?

[GreenKoopa]

Sorry, I'm not sure what you mean by "compute (black) connected components". The only things that come to my mind are a flood fill tool, a fuzzy select magic wand tool, or morphology, and then I'm not sure what you are doing with them.

[snibgo]

I'm also not sure what you mean. Perhaps, "Given an image that contains only black and white pixels, how do I count the number of components, where a component is a number of connected black pixels?"

The answer is a rather horrible and slow script.

The process is:

1. Find a black pixel.
2. Flood-fill from that point, turning black into white.
3. Increment a counter.
4. Repeat from 1 until there are no more black pixels.

[izavorin]

Sorry, I'm not sure what you mean by "compute (black) connected components". The only things that come to my mind are a flood fill tool, a fuzzy select magic wand tool, or morphology, and then I'm not sure what you are doing with them.

Given a black and white image, produce a mask "image" of the same size (or something equivalent) that would have labeled pixels according to whether they belong to a specific black connected component based on 4- or 8-pixel neighborhoods. If you are familiar with the Image Processing toolbox of Matlab, it would be what

Code: Select all

L = bwlabel(~im);

would produce, where

Code: Select all

im

is a black and white image.

[GreenKoopa]

Producing a mask image is a common technique. The only IM commands I know that work with connected components are -floodfill and -draw floodfill. These can be used to create a mask in a round about way, or used more directly if that makes sense. I don't know if they use 4- or 8-pixel neighborhoods, but that would be easy to test. I can't recall seeing an example, but I'll look.

Note that internally IM doens't work in B&W or even grayscale. All images are full RGB (at least until version 7). You can achieve great results all the same, but learning that helped me think about commands more clearly.

[GreenKoopa]

Given a black and white image, produce a mask "image" of the same size that would have labeled pixels according to whether they belong to a specific black connected component based on 4- or 8-pixel neighborhoods.

Here is an example:

Code: Select all

: Create test image
convert -size 100x60 xc:white -fill black -draw "rectangle 30,10 70,30 rectangle 10,30 30,50 rectangle 71,31 90,50 rectangle 40,40, 60,50" image.png

: Optioin 1, using -floodfill
convert image.png -fill green -floodfill +50+20 black filled_1.png
convert filled_1.png -fill black +opaque green -fill white -opaque green mask_1.png

: Option 2, using -draw
convert image.png -fill green -draw "color 50,20 floodfill" filled_2.png
convert filled_2.png -fill black +opaque green -fill white -draw "color 50,20 replace" mask_2.png

IM appears to use 4-pixel neighborhoods for floodfill.

--- EDIT ---
Changed -opaque white to +opaque green to make the example work beyond the 1-bit B&W case.

An alternative option to creating a mask is to introduce an alpha channel. For finding connected areas you are stuck with the first step of floodfill, but there are many options for the next step.

[izavorin]

Here is an example:

Thanks, I'll experiment with this!

[anthony]

You can also do multi seed point flood fills with user defined neighbourhoods using Conditional Dilation
Conditional Dilation is actually what 'flood fill' is doing, just using a fast more specific algorithm.

See Conditional Dilation
http://www.imagemagick.org/Usage/morpho ... onditional