Detecting/Removing Vertical Scanlines from Scans

[Tex2002ans]

Does anyone have a method for detecting/removing lines in scans WITHIN text?

I think this would really help towards cleaning up PDF scans from Archive.org and elsewhere.

The Problem

(These sample images were cleaned up manually: Before = ScanTailor, After = Manual Cleanup.)

"Easy" Lines

These type of lines occur where the line barely intersects with text:



"Hard" Lines

The worst though, are lines which go through the middle of text.

What I would ultimately aim for is Before/After:



Potential Solutions

This is some of the research I've done so far:

#1: Hough Lines (ImageMagick)

To detect lines within images, I ran across "Hough-Line Detector" in the ImageMagick Manual.

This site+script shows how Hough Lines can detect lines within images (photographs):

http://www.fmwconcepts.com/imagemagick/houghlines/
https://www.imagemagick.org/discours...ic.php?t=25476

Towards the bottom of the fmwconcepts link, it shows how to detect the horizon (or a fence) by inverting the image.

Maybe there's a way to use inversion to detect vertical lines easier:



Related: And here's a different forum post using Hough Lines to detect page boundaries:

http://www.imagemagick.org/discourse...ic.php?t=31321

(Although I find ScanTailor already does a decent enough job at page boundary removal.)

#2: ImageMagick (Removing Lines Outside of Text)

Some of these examples show removing horizontal/vertical lines... but these are straight lines with no breaking in between:

Remove Vertical Lines for Pre OCR (Tesseract)

These show removing table borders from an image:

https://stackoverflow.com/questions/...s-programmatic

https://stackoverflow.com/questions/...ize-from-image

#3: Baseline Detection + Vertical Line Detection

Perhaps there is a way to combine "baseline detection", with a "vertical line detector"... to at least delete MOST of the vertical line automatically.

Even this would be a huge help over complete manual cleanup.

(I roughly put these images together in GIMP. Still haven't figured out how to get ImageMagick to do this.)

Step 1. Detect Baselines + Expand:



Step 2. Detect Vertical Line (potentially use Hough Lines?):



Step 3. Automatically remove vertical line that DOESN'T overlap the baselines:



While not as good as manual cleanup, this would make the text much more readable + OCR-friendly. See Original + "Automatic" + Manual pages side-by-side:



Related: Here's ImageMagick being used to add Line Numbers to an image. Perhaps something from here can help detect baselines.

* * * *

Hoping everyone can put their minds together and come up with some solution.

[BetterRed]

Have a look for imaging forensic tools. Twenty years ago some of them could detect inconsistent and unusual edges, can't recall if they could remove them - probably. Back then they cost four figures and ran on SG and similar kit.

BR

[Tex2002ans]

Originally Posted by BetterRed

Have a look for imaging forensic tools.

Hmmm... that may be another angle to research.

I know a lot of times they "average" pixel colors of entire rows/columns to get semi-unique fingerprints. Perhaps something like that could be used to detect lines too.

* * *

Tonight I was dabbling a bit more with Hough Lines.

I had quite a bit of success locating the line through the text.

ImageMagick Hough Lines

Original Image:



Step 1: Inverse the scan using ImageMagick's canny (see fmwconcepts link in Post #1):

convert test.png -canny 0x1+10%+40% test_inverse.png

Step 2: Then calculate Hough Lines:

From testing, on this specific book, I found a threshold between 500-700 worked:

Spoiler:

convert -hough-lines 5x5+500 -fill red -transparent white test_inverse.png test_lines_500.png
convert -hough-lines 5x5+550 -fill red -transparent white test_inverse.png test_lines_550.png
convert -hough-lines 5x5+600 -fill red -transparent white test_inverse.png test_lines_600.png
convert -hough-lines 5x5+650 -fill red -transparent white test_inverse.png test_lines_650.png
convert -hough-lines 5x5+700 -fill red -transparent white test_inverse.png test_lines_700.png

The higher the threshold, the more "false positives" disappeared.

Step 3: Overlay Hough Lines with image:

Spoiler:

convert test.png ( test_lines_500.png ) -compose over -composite test4_composite_500.png
convert test.png ( test_lines_550.png ) -compose over -composite test4_composite_550.png
convert test.png ( test_lines_600.png ) -compose over -composite test4_composite_600.png
convert test.png ( test_lines_650.png ) -compose over -composite test4_composite_650.png
convert test.png ( test_lines_700.png ) -compose over -composite test4_composite_700.png

Here's the same steps with another page:



* * *

Side Note: To see what a Hough Line calculation is actually doing, I found this part of the video did a decent job explaining it visually:

https://youtu.be/4zHbI-fFIlI?t=219

It goes row-by-row detecting each white pixel, then spins a line in a 360. Plotting this leads to points of various strength (which tells you probable locations + angles of lines).

[Turtle91]

They have some image correction/ocr software over at diybookscanner.org that may be of help...it's some pretty amazing stuff.

[Tex2002ans]

Originally Posted by Turtle91

They have some image correction/ocr software over at diybookscanner.org that may be of help...it's some pretty amazing stuff.

I'll definitely have to check out their forums again. Haven't visited them in a few years, perhaps someone tackled a similar issue.

I've been using Scan Tailor Advanced lately, and that's initially how I cleaned up those "Original" images. It does a fantastic job cropping and correcting the distortions. (And runs so much faster/better than the original Scan Tailor OR Scan Tailor Enhanced.)

... but this line-through-text issue has niggling away at me for a while now. Usually it's not so bad, and I can manually clean it, but this particular book scan had a vertical line on EVERY even page.