How to delete this background?

[devtra]

Hi,

I have the following images:

source: https://imgur.com/58lyq2W

source: https://imgur.com/sNdyHAn
I was unsuccessful to delete all the noisy background and let only the text. I have used this great script by Fred, textcleaner and it gave me sometimes good results sometimes no. So I thought I could delete all the colors in the image except Black and it's variants.
So what I did was

Code: Select all

convert image.jpg -fill white -fuzz 40% +opaque "#000000" removed.png 

and I start playing with the fuzz parameter. The problem was that the text are not detected as black, so I thought of converting all the variants of black and grey to only black then proceed with the removal of other colors.

I tried also to binarize the image and clean it as follows:

Code: Select all

localthresh -m 3 -r 25 -b 20 -n yes image.png output.png
textcleaner -b white -g -e normalize -f 35 -o 6 -t 50 -S 200 -T output.png binarize.png
denoise -f 2 -s "20x20+203+152" binarize.png binarize.png

But the background still appears as like so:


source: https://imgur.com/HZsYI7G

source: https://imgur.com/jA4R6vD

Sometimes even worse.
Is there anyway I could do that, I am still a newbie and searching on google didn't give me much.
Or if there is another procedure to follow to get better results it would be appreciated.

Thanks you.
Regards!

[devtra]

Update 1:
I found out that doing a threshold before using the textcleaner script affects the performance of the image, this is why the noise in the background appears.
So what I did is applying textcleaner directly with high threshold (-t parameter) and big filter size (-f parameter) and offset size (-o parameter).

Code: Select all

./textcleaner -b white -e normalize -f 41 -o 10 -t 90 -S 200 -T image.jpg binarize.png

Th result was outstanding with High quality images.

But I have faced another problem: now there are holes in my letters, small ones though , but they affect badly the performance of the tesseract. So, is there any way to fill those holes with black color, like an average of nearing pixels, or a predetermined kernel: diamong, rectangle, etc?

Have a nice day.

[fmw42]

Please show input and output examples of your latest command with text cleaner

[devtra]

Image input:

source: https://imgur.com/OLeaep7

Image output:
source: https://imgur.com/whKg0LE


You see what I mean, look at the characters in the top left: N J5020
They are not clear enough. If you zoom in, you will see white holes.
Although the background was removed, it took some from the chars.

Hops that clears the problem.

[fmw42]

That is not an easy thing to solve. Your scan needs to be a higher resolution and not blurry. It is a bit blurry because it is in perspective and parts are out of focus. I do not know how to solve it.

[devtra]

I see. Thank you for your concern.
The problem is not with the text itself, I am more concerned with the background.
Is the only way to remove the background the one I used, isn't there any other consistent method.
I'd love to hear your opinion.
Regards!

[fmw42]

Background removal is very difficult except for a nearly constant color background. Any texture to the background makes it very hard if not impossible to remove adequately. Your dark background has bright spots in it that make it hard to remove since the bright spots are nearly like the color of your text background. However, you could try using canny edges and hough line detection to get the region of your license identified by the 4 corner intersections of the lines. Then make a mask of that and then use that to remove your background. See http://www.imagemagick.org/discourse-se ... =4&t=25476. But doing so will require a bit of scripting to find the intersections and draw the bounding box.

[devtra]

I can already detect all the bounding boxes that I want inside the image. I used openCV.
I can detect the outer box also. See the image below that I have generated.

source: https://imgur.com/h04fU1c