Any suggestions on pre-processing to improve accuracy?

[Traun Leyden]

I'm wondering how I can get better results with Tesseract.  

Here are a few images I've been testing with + results:

Image: 

Actual OCR text: VCZZSWE

Expected OCR text: VC22500E

Image:  

Actual OCR text: ViZZSWE DRIVEWAY

Expected OCR text: VC22500E DRIVEWAY

Any tips on doing pre-processing on the images to improve the recognition?

The code I'm using to call tesseract (via go-tesseract) is here: 

  https://github.com/tleyden/open-ocr/blob/master/tesseract_engine.go#L49-L53

Version: I'm using the tesseract-ocr-eng package from Debian Jessie, which looks to be version: 3.02-2  (the full build script is available in this Dockerfile)

[Nick White]

Hi Traun,

> Any tips on doing pre-processing on the images to improve the
> recognition?

The place to start would be here:

https://code.google.com/p/tesseract-ocr/wiki/ImproveQuality

Nick

[Traun Leyden]

Thanks, this is really useful.  (and shame on me for not RTFM'ing a bit more first)

That document mentions to make sure the orientation/skew is straight, but does not give any hints on how to actually do this in an automated fashion.  Any tips?

[Robert Komar]

On Fri, 20 Jun 2014, Traun Leyden wrote:

> Thanks, this is really useful. (and shame on me for not
> RTFM'ing a bit more first)
> That document mentions to make sure the orientation/skew
> is straight, but does not give any hints on how to
> actually do this in an automated fashion. Any tips?

You can use imagemagick's "convert" utility to deskew
images. For example:

> convert <skewed_image> -deskew 40 <deskewed_image>

It works pretty well for text-only images. Embedded
images within the text tend to mess it up, though.

Rob Komar

[Traun Leyden]

Thanks, I will definitely use that.

One more thing that document should have is a mention of Stroke Width Transform to improve OCR recognition on images that have a lot of non-text content.

Here's an example of SWT in action.

--
You received this message because you are subscribed to a topic in the Google Groups "tesseract-ocr" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/tesseract-ocr/tjY0jZsopwA/unsubscribe.
To unsubscribe from this group and all its topics, send an email to tesseract-ocr+unsubscribe@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at http://groups.google.com/group/tesseract-ocr.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/alpine.LNX.2.02.1406221104320.7820%40robpc4.home.org.

For more options, visit https://groups.google.com/d/optout.

[Nick White]

On Mon, Jun 23, 2014 at 08:32:52AM -0700, Traun Leyden wrote:
> One more thing that document should have is a mention of Stroke Width Transform
> to improve OCR recognition on images that have a lot of non-text content.

Oh cool, that looks great! I definitely will add that to the wiki
page soon, thanks alot for pointing me to it.

Nick

[Traun Leyden]

Btw I original came across it via Project Naptha, which is using Tesseract for it's underlying OCR.

--
You received this message because you are subscribed to a topic in the Google Groups "tesseract-ocr" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/tesseract-ocr/tjY0jZsopwA/unsubscribe.

To unsubscribe from this group and all its topics, send an email to tesseract-oc...@googlegroups.com.

To post to this group, send email to tesser...@googlegroups.com.
Visit this group at http://groups.google.com/group/tesseract-ocr.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/20140626200201.GA30921%40manta.lan.