[Tex2002ans]

Originally Posted by roger64

Yes I tried Scan Tailor advanced (but it's last year version now...) and I was disappointed: too complex for me, unstable.

Looks exactly the same as Scan Tailor to me, just has a few optional tabs/buttons in some steps.

But maybe the Linux version is less stable. The Windows version for me has been solid as a rock (and much faster than all the previous ones, since it's heavily multi-threaded).

Originally Posted by roger64

Some PDF though are beyond repair: (but it's exception) https://gallica.bnf.fr/ark:/12148/bp...811.texteImage

Yuck, looks about as bad as mine... but yours can be solved!

I was able to follow most of the steps here:

Removing noise from scanned text document

Step 1

Get the PDF into PNGs:

convert -density 300 input.pdf output.png

Since that PDF is awful, and has enormous whitespace around it, I would suggest trimming:

convert -density 300 input.pdf -trim output.png

that would focus more on the text itself.



Alternate #1: You could also use the magick.exe command:

magick.exe -density 300 input.pdf -trim .\output-%d.png

ImageMagick ran out of memory on my end, so if you want to convert the PDF in pieces, you can adjust the [0-30] to fit whatever page numbers you want to export:

magick.exe -density 300 input.pdf[0-30] -trim .\output-%d.png

Side Note: Not sure why it's only cropping vertically, there's probably another method to crop the left/right whitespace too. It would probably speed up the later steps too.

Step 2

Now, I followed much of that forum post above.

convert output.png -connected-components 4 -threshold 0 -negate output-negate.png

Step 3

It seems like area-threshold looks for "chunks of pixels that are X pixels or less".

I tested with area-threshold=30:

convert output.png -define connected-components:area-threshold=30 -connected-components 4 -threshold 0 -negate output-cc30.png

but I found that this PDF needed more. So I adjusted by 10s all the way up to 80:

convert output.png -define connected-components:area-threshold=80 -connected-components 4 -threshold 0 -negate output-cc80.png

Step 4

Then I was able to take image from Step 1 + Step 3 and create a diff:

convert output.png output-cc80.png -compose minus -composite output-diff.png

Step 5

Then use the images from Step 1 + Step 4 to remove:

convert output.png ( -clone 0 -fill white -colorize 100% ) output-diff.png -compose over -composite output-diff-composite.png

Here's the Original (Step 1) + Diff (Step 3) + Cleaned (Step 5):



Finalized

Here's a few more before/after pages out of the book:




I attached a ZIP with Windows .bat files that batch convert the images using these steps. It's a giant mess, and it does create a lot of blank/duplicate images, but it chugs through everything eventually. I already spent hours writing this tutorial up, and don't feel like debugging the rest.

But hopefully that'll get you much cleaner input into Scan Tailor + better OCR.