[Chicago] OpenALPR - Transformations

[Jason Wirth]

Could you be a little more specific in the problem you have; I looked at
the sample image on the site and tried a couple random images on the demo
and it detected a skewed plate so I'm not sure what's not working. Is it
that yours is highly skewed while theirs only detects moderately skewed
images, or it cannot read the numbers correctly?

Also, what do you mean by "sometimes that it doesn't read it correctly"?
What's the overall accuracy?

In many cases you can automatically create a huge training data set by
taking a good image then applying various tilt, skew, and reshaping
transformations. Perhaps you can train something yourself.






-- 
    Jason Wirth
    wirth.jason at gmail.com

On Fri, Nov 4, 2016 at 2:38 PM, Robare, Phillip (TEKSystems) <
proba at allstate.com> wrote:

> Some links you may find useful below.  I am surprised ALPR doesn't do this
> already.
>
> The problems reading license plates by humans have resulted in deaths (due
> to hits on mis-entered license numbers in the context of traffic stops).  I
> hope you are getting a better feeling of how we need to be careful with the
> tasks given to our new robotic overlords.
>
> An interesting post from a couple years ago is "How to Build a Kick-Ass
> Mobile Document Scanner in Just 5 Minutes" (http://www.pyimagesearch.com/
> 2014/09/01/build-kick-ass-mobile-document-scanner-just-5-minutes/) .  It
> covers sharpening the image, doing a perspective transformation and getting
> the document ready for OCR (but not the OCR step).
>
> The book "Automate The Boring Stuff With Python" has a chapter on
> "Manipulating Images" that is online and covers how to do the manipulations
> with PIL (https://automatetheboringstuff.com/chapter17/).  This would be
> a good start.  If you are working on your project with Python 3 (if not,
> why not?) I believe Pillow has the same API.
>
> There have been a number of github projects in the space of preparing
> scanned documents that I have seen over the years.  Unfortunately I can't
> find the best ones I remember so here are ones I found today:
> Scantailor (https://github.com/scantailor/scantailor) is a large C++
> project that appears to be very complete.  ImproveQuality (
> https://github.com/tesseract-ocr/tesseract/wiki/ImproveQuality) is also a
> cpp project meant for text scanning.
>
> A mathematically interesting approach to deskewing is documented in
> http://www.ijstr.org/final-print/dec2013/An-Integrated-
> Skew-Detection-And-Correction-Using-Fast-Fourier-Transform-And-Dct.pdf
> where a Fast Fourier Transform is used to determine the skew without
> pre-analyzing the image to pull out text lines.
>
>
> Phil Robare
>
>
> -----Original Message-----
> From: Chicago [mailto:chicago-bounces+proba=allstate.com at python.org] On
> Behalf Of Chris Vinzons
> Sent: Thursday, November 03, 2016 2:52 PM
> To: chicago at python.org
> Subject: [Chicago] OpenALPR - Transformations
>
> Hi,
>
> I'm using OpenALPR to read still license plates, but the thing is that my
> sometimes that it doesn't read it correctly. I think this is because of the
> license plate is tilted. Is there a way to untilt it or is there some kind
> of training data I could do in python?
>
>
> Thanks
>
> Chris V
>
> _______________________________________________
> Chicago mailing list
> Chicago at python.org
> https://urldefense.proofpoint.com/v2/url?u=https-3A__mail.
> python.org_mailman_listinfo_chicago&d=DQICAg&c=gtIjdLs6LnStUpy9cTOW9w&r=
> VXIryE9UwJGlNMLzgMzDT4_t2NMrZf6alSphHwSEwC0&m=aGirqomLg7oSQWy-
> OKQqUWrYZDOqsWUYn7NEEJF2wo0&s=IbFXGkHqKjSJk4LX4ewkoVwyajgo6U
> OWIoSznB8I7ps&e=
> _______________________________________________
> Chicago mailing list
> Chicago at python.org
> https://mail.python.org/mailman/listinfo/chicago
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/chicago/attachments/20161104/7323c1c1/attachment.html>