dewarping scanned images

[Romeyke, Andreas]

Hello,

We need to develop an image dewarping function. We have a working
function to evaluate the distortion. What we need is a transfer
function which maps P(x,y) to P'(x',y') based on a model of typical
distortions of book scans.

We have experimented with following functions (polynom based), but they
need too many parameters:

ax + by + c fx + gy + h
x' = -----------, y' = -----------
1 + dx + ey 1 + dx + ey

If we have a look to typical book form, we think something similar to
this would be better:

b -cx
x' = ax e + d, with a~1, b~0.5, c~1, d~0

Do you have another suggestions? Do you have ideas to map y -> y'?

Thanks for your help...

Bye Andreas
--
Andreas Romeyke
- Abteilung Blindenschrift -
Deutsche Zentralbücherei für Blinde zu Leipzig (DZB)
Gustav-Adolf-Straße 7, 04105 Leipzig
Tel: +49 341 7113-..., Fax: +49 341 7113-125
Internet: www.dzb.de
E-Mail: andreas...@dzb.de

--
--
Diese e-mail wurde auf Spam und Viren mit Astaro Security Gateway geprueft.

[Tom]

That depends on many factors. Are you talking about dewarping of
camera-captured images, or dewarping of flatbed scanned pages? And
what purpose do you want to dewarp the images for?

Dewarping of camera-captured images involves projective transforms and
taking advantage of the fact that paper is physically constrained and
can curl only in certain ways. In addition, camera lenses may be
distorting the image as well. We're developing software for doing
this as part of the DECAPOD project, but it's still going to take a
while until that gets released.

Dewarping scanned pages involves compensating for periodic slippage
and excentricity in the scanner transport. That's a completely
different model.

Generally, no simple global model is going to work well for either
situation.

If this is just for OCR, we're currently integrating layout analysis
into OCRopus that can deal with some degree of curvature. With that,
you should be able to perform OCR on lightly distorted pages directly
without dewarping.

Tom

> E-Mail: andreas.rome...@dzb.de

>
> --
> --
> Diese e-mail wurde auf Spam und Viren mit Astaro Security Gateway geprueft.
>

>  signature.asc
> < 1KViewDownload

--
You received this message because you are subscribed to the Google Groups "ocropus" group.
To post to this group, send email to ocr...@googlegroups.com.
To unsubscribe from this group, send email to ocropus+u...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/ocropus?hl=en.

[Romeyke, Andreas]

Hello,

Am Donnerstag, den 13.05.2010, 12:34 -0700 schrieb Tom:
> That depends on many factors. Are you talking about dewarping of
> camera-captured images, or dewarping of flatbed scanned pages? And
> what purpose do you want to dewarp the images for?

Both, or in detail we are using camera based flatbed scanners where the
book is pressed against a glass plate and scanned by a camera on top.

Also we used classical flatbed scanners.

> Generally, no simple global model is going to work well for either
> situation.
>

We know and that is the reason to ask you :)

> If this is just for OCR, we're currently integrating layout analysis
> into OCRopus that can deal with some degree of curvature. With that,
> you should be able to perform OCR on lightly distorted pages directly
> without dewarping.

We have developed a segmenter which could handle this. The problem is
on OCR side. We are planning to use a lot of different engines and we
think we should not trust them to do the job well. ;-)

Bye Andreas
--
Andreas Romeyke
- Abteilung Blindenschrift -
Deutsche Zentralbücherei für Blinde zu Leipzig (DZB)
Gustav-Adolf-Straße 7, 04105 Leipzig
Tel: +49 341 7113-..., Fax: +49 341 7113-125
Internet: www.dzb.de

E-Mail: andreas...@dzb.de