Re: [tesseract-ocr] '33' recognized correctly, '3' not recognized at all...

[Shree Devi Kumar]

ubuntu@tesseract-ocr:~/TEST$ tesseract twonumbers.png - --psm 6 --tessdata-dir ~/tessdata --oem 1
2 127

a 15

7 56

7 58

9 58

19 65
24 91
3375
ubuntu@tesseract-ocr:~/TEST$ tesseract twonumbers.png - --psm 6 --tessdata-dir ~/tessdata_best --oem 1
2 127

a 15

7 56

7 58

9 58

19 65
24 91
3375
ubuntu@tesseract-ocr:~/TEST$ tesseract twonumbers.png - --psm 6 --tessdata-dir ~/tessdata_fast --oem 1
2 127

4 15

7 56

7 58

9 58

19 65
24 «(91
33 «75

On Sat, Aug 31, 2019 at 9:54 PM Jack <otak...@gmail.com> wrote:

I have a weird niche project here, essentially I have about 4,000 images, each with 2 numbers between 0 and 127.

I've tweaked the images in a million different ways and I can't get tesseract to recognized individual numbers, with the exception of 2, all other 1 digit numbers are not recognized.

Also, for some reason if I use tesseract directly I get way worse results, whereas if I convert to pdf first and use ocrmypdf, which apparently uses tesseract, I get WAY better results, which I don't understand.

The font is very straight-forward I think, so I'm not sure if training would be helpful, but I'm open to the idea if needed.

Here are the sample images I'm using for testing, before and after I modified them:

Before: https://imgur.com/a/PhjWXXK

After: https://imgur.com/a/sCRE67S

Okay some of them failed to upload but that's the gist.

Thanks,

Jack

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/7be5ed42-df44-4530-b7a2-0d0fa340918e%40googlegroups.com.

--

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

[Jack]

Thank you for replying, that was very helpful.

I've now tried tessdata_best and tessdata_fast trained data found on the tesseract github, which has drastically improved my results, but still not as accurate as yours.

Here are my outputs:

tesseract listpng output2 --psm 6 --tessdata-dir ~/tessdata/tessdata_best --oem 1

3 70
2 127
4 15
7 96
7 98
9 B58
9 65
19 695
29 91
33 75

tesseract listpng output_fast --psm 6 --tessdata-dir ~/tessdata/tessdata_fast --oem 1

3 70

2 127
4 15
7 56
7 58

9 #58
9 #65
19 ~=665
24 #691
33 #675

[Jack]

Thank you very much for your reply, that was very helpful, I think that should do the trick.

To unsubscribe from this group and stop receiving emails from it, send an email to tesser...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/7be5ed42-df44-4530-b7a2-0d0fa340918e%40googlegroups.com.

[Shree Devi Kumar]

I am using the latest code from master branch. 

I would expect same result with same image and same traineddata files.

--

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 view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/934d89f8-a455-4787-8d8d-8986cc615059%40googlegroups.com.

[Shree Devi Kumar]

Well, I just took a screenshot of your images from the link since I could not figure out how to get individual images. The only doctoring was to save it at 300 dpi in irfanview.

On Sun, 1 Sep 2019, 08:27 Jack, <otak...@gmail.com> wrote:

Ah, now I see it has something to do with the way you doctored the images, I get the same output as you did when I ran your pic through. So what's the secret?

On Saturday, August 31, 2019 at 11:24:23 AM UTC-5, Jack wrote:

I have a weird niche project here, essentially I have about 4,000 images, each with 2 numbers between 0 and 127.

I've tweaked the images in a million different ways and I can't get tesseract to recognized individual numbers, with the exception of 2, all other 1 digit numbers are not recognized.

Also, for some reason if I use tesseract directly I get way worse results, whereas if I convert to pdf first and use ocrmypdf, which apparently uses tesseract, I get WAY better results, which I don't understand.

The font is very straight-forward I think, so I'm not sure if training would be helpful, but I'm open to the idea if needed.

Here are the sample images I'm using for testing, before and after I modified them:

Before: https://imgur.com/a/PhjWXXK

After: https://imgur.com/a/sCRE67S

Okay some of them failed to upload but that's the gist.

Thanks,

Jack

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/bca34cd1-5092-4601-bd98-ee3e3d2aa5ab%40googlegroups.com.