Dot Matrix Fonts and Tesseract's Connected Component Analysis

[ameera...@gmail.com]

I am trying to fine-tune Tesseract for dot-matrix fonts such as that in the picture below.  When the dots are closely spaced together and touch, Tesseract can more or less handle the dot-matrix font with some fine-tuning and image processing.  However, when the dots do not touch, as in the picture below, Tesseract struggles.

I read in An Overview of the Tesseract OCR Engine that the first step in Tesseract's processing pipeline is a connected component analysis (second paragraph of Section 2).  Since the letters in a dot-matrix font do not form connected components, I am wondering if Tesseract's connected component analysis may be one reason that Tesseract struggles on the image below.  

Is there a command to see how Tesseract performs connected component analysis on this image?  

[Shree Devi Kumar]

> I read in An Overview of the Tesseract OCR Engine that the first step in Tesseract's processing pipeline is a connected component analysis (second paragraph of Section 2).

That applies to base/legacy tesseract (3.0x). Tesseract4 is neural net/LSTM based though it still supports the legacy models.

I tried some finetuning using dot matrix fonts. While the image you gave is not recognized correctly, if split in separate lines the results are better, but only with --psm 8 (single word).

/home/ubuntu/TEST/ex_20_1 dotslayer
LOT#S18FO70

/home/ubuntu/TEST/ex_20_2 dotslayer
EXP:09/2020

The traineddata file is attached.

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at https://groups.google.com/group/tesseract-ocr.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/fbbc3452-62f5-4c34-bd9c-72fa3a52c97c%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

____________________________________________________________
भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com

[ameera...@gmail.com]

Hi Shree,

Thanks for sending these images and the traineddata file.  I confirmed that they worked.  Would you please tell me a little bit more about what kind of image processing you used to make the .png images and how you created your traineddata file using fine-tuning?

Thank you,

Ameera

[Shree Devi Kumar]

Hi Ameera,

Please do check with other images too as I tested with only one image that you sent.

I had initially tried fine tuning (impact and plus) but those were not giving accurate results for 2nd line.

Then I tried replace the top layer, using new training text all in UPPER case, with many lines in the same format as the image u sent. I used just a couple of fonts that looked similar to the image.

Regarding the image, I tested different versions by changing it interactively in irfanview. Mainly, straighten the image, convert to black and white , resize to half and then half again. I haven't tested the new traineddata with the original image.

I will email you the training text and fonts used, if you want.

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at https://groups.google.com/group/tesseract-ocr.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/7a30ee84-cae8-406f-82e1-ca7767e40f20%40googlegroups.com.

[Shree Devi Kumar]

Also changed image to 300 dpi and used --dpi 300.

[ameera...@gmail.com]

Hi Shree,

Thanks for the files!  That's interesting that you tried replacing the top layer.  I haven't tried that yet.  How many iterations did you use?

I was thinking today that it is difficult to create a single strong learner with tesseract because training from scratch requires so much data.  However, with fine-tuning, it is easy to create a lot of weak learners.  I am wondering if you know of any successes of an ensemble model with tesseract.

Thanks again,

Ameera

On Friday, March 22, 2019 at 12:11:11 AM UTC-7, ameera...@gmail.com wrote:

[Shree Devi Kumar]

That's interesting that you tried replacing the top layer.  I haven't tried that yet.  How many iterations did you use?

In this case the unicharset was limited to UPPERCASE letters, 0-9 numbers , : and /.

I used a training_text which followed the pattern of the image - lines starting with LOT# and EXP: and using similar pattern.

I used 2 fonts which were very similar to the image.

So this was narrowly focussed on single use and only 2000 iterations were needed with tessdata_best/eng to get error rate down to 0.2 or so.

The # of iterations for plus training were also similar but they did not give same accuracy (also, the traineddata file size is much smaller using this method).

[Balaji Gurunathan]

Hi,

I've a similar requirement to read dot-matrix fonts but I'm not sure where to begin this from since I'm new to Tesseract. Could you please share references/guide.

Thanks.

[ameera3]

Hi Balaji,

You may find this GitHub repository useful, https://github.com/ameera3/OCR_Expiration_Date

Best wishes,

ameera3