search

lessthanoptimal / BoofCV

Fast computer vision library for SFM, calibration, fiducials, tracking, image processing, and more.
http://boofcv.org
1.08k stars 260 forks source link

Feature Request: Add wolf binarization #124

Closed sandreas closed 3 years ago

sandreas commented 5 years ago

Hey,

i would appreciate, if you would add wolf binarization to boofcv. Seems to be a very nice little algorithm with impressive results (much better than sauvola).

Sample Code in C: https://github.com/chriswolfvision/local_adaptive_binarization

What do you think?

lessthanoptimal commented 5 years ago

Skimming the paper it looks like it should be straight forward and has a decent number of citations. It won't make it into the up coming stable release but it could make it into the next.

sandreas commented 5 years ago

Nice :-) Thank you.

thhart commented 5 years ago

Hi, I don't want to be this septic but are you sure it is this superior related to Sauvola? On the GitHub page the example is really bad. From my experience Sauvola and Otsu (slow) are extremely good with a good parameter set.

sandreas commented 5 years ago

Hi, I don't want to be this septic

Its ok :-) In a discussion, nothing is septic. Especially because I'm not a professional in this topic.

are you sure it is this superior related to Sauvola

Not 100% sure, but all citations i read were superior in many cases.

On the GitHub page the example is really bad

Bad in what way? Not representative or bad result? If you think its a bad result, are you sure you did not mix it up?

In my opinion, the fourth result is the best one for representing the ground truth...

lessthanoptimal commented 5 years ago

Tangentially related, these thresholding algorithms are now threaded. About 1X faster for each core in your computer if concurrency is turned on.

thhart commented 5 years ago

Which algorithms in which version?

lessthanoptimal commented 5 years ago

The latest SNAPSHOT, which is a candidate for the next stable. Just turned on threading by default. I think all thresholding algorithms have at least part of the code parallelized now.

sandreas commented 5 years ago

this is great news, thank you :-)

lessthanoptimal commented 4 years ago

https://perso.liris.cnrs.fr/cwolf/papers/icpr2002v.pdf

Just saving a reference to the paper. It's got enough citations and source code that it's worth taking a closer look at it

lessthanoptimal commented 3 years ago

Resurrecting this ticket. Just added it to the "official" list of features to be added in this up coming release. Only been about two years, still need it @sandreas ? ;-)

sandreas commented 3 years ago

@lessthanoptimal It's not that I need it, but let me ask if there is a superior thresholding algorithm implemented and how to use it? ;-)

lessthanoptimal commented 3 years ago

https://github.com/lessthanoptimal/BoofCV/pull/243

Turns out this was very easy. Niblack, Sauvola, and Wolf are 99% the same code, the inner most function is different. Wolf and Sauvola look very similar unless the image is over saturated. There are also two different versions of Sauvola out there. The one in BoofCV is based on the journal article. No actual change in behavior though.

lessthanoptimal commented 3 years ago

I think all of the more famous "classical" thresholding algorithms have been added. I wonder how difficult adding basic OCR would be.

sandreas commented 3 years ago

I already experimented with some "basic" OCR - this should not be underestimated. The process of pure OCR may already be complex enough because of the big variety of fonts and languages out there - not talking about RTL languages, but the preprocessing steps (thresholding, deskewing, etc.) are also very important (see https://towardsdatascience.com/pre-processing-in-ocr-fc231c6035a7).

I once found a paper about line detection on geographical maps (bended lines, different colors) that looked very promising but I did not have the time to try it out with BoofCV. Unfortunately I also did not find it again nowadays...

But let me say that I would love to see a more professional BoofCV approach to OCR with LTSM or other AI techniques, if you find the time :-)

lessthanoptimal commented 3 years ago

I'm thinking of something very simple to start with. Targeting 5 or so very similar blocky "industrial" fonts in a non cluttered image. Just alpha numeric English characters with clear separation. i would prefer to not use any DL approaches since those will be very slow on a CPU and are probably handled better by a dedicated library. If you know anything that would fit the bill let me know!

sandreas commented 3 years ago

Maybe this: https://www.cs.bgu.ac.il/~berat/papers/mdpi2020Learningfree.pdf

https://www.cs.bgu.ac.il/~berat/