Tesseract not giving the desired output

[Vidur Malhotra]

I tried running the tesseract on the attached image. But not getting the desired output. My sample code:

import PIL

from PIL import Image

import pytesseract

text = pytesseract.image_to_string(Image.open('test3.jpg'), lang='eng')

print(text)

[sett...@gmail.com]

Not a solution, but the image needs to be transformed into grayscaleetc, (using Open CV) since OCR works best with grayed images and images which have size of 300 dpi

[Zdenko Podobny]

That is not true: you do not need to transform image to grayscale. Any image is at the end binarized (if input image is not binarized) by tesseract (Otsu).

BUT: preprocessing image (e.g. custom binarization) will help. See https://github.com/tesseract-ocr/tesseract/wiki/ImproveQuality

Zdenko

st 30. 1. 2019 o 7:36 <sett...@gmail.com> napísal(a):

--
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/43a5a754-4227-43b6-aec1-0261403b2029%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Lorenzo Bolzani]

Zdenko, are you 100% sure that the image is binarized before being fed to the neural network? It looks like a big waste of information to me.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAJbzG8zsShhi_LsUXkCWoj8uxWxROkTL7G4RpGwzBEVm1EweTA%40mail.gmail.com.

[Zdenko Podobny]

try:

 tesseract image - get.image

which calls GetThresholdedImage()

Zdenko

st 30. 1. 2019 o 11:17 Lorenzo Bolzani <l.bo...@gmail.com> napísal(a):

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAMgOLLxVGXt527-wkSfGegMtOjMU2LT0rz_H%3Dp8kQZ13CCE1ag%40mail.gmail.com.

[Lorenzo Bolzani]

I suppose this means that the image is always binarized, is this correct?

Is there any way to avoid it?

Does this binarization happens by default during training too?

I fine tuned a few models using grayscale images. Do you thing the neural network received binary black/white pixels or the gray ones?

Thanks, bye

Lorenzo

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAJbzG8zse4y474Uz7w0pJEuyDesRgD6fuQu_Y0cMDzGH4Ux7JA%40mail.gmail.com.

[Zdenko Podobny]

see inline comments.

st 30. 1. 2019 o 15:17 Lorenzo Bolzani <l.bo...@gmail.com> napísal(a):

I suppose this means that the image is always binarized, is this correct?

Yes 

Is there any way to avoid it?

Why? IMO OCR engines are running on binarized images see e.g. https://www.abbyy.com/en-eu/ocr-sdk/key-features/image-processing/

Does this binarization happens by default during training too?

I do not know. I did not have time to play with training in v 4.0 

I fine tuned a few models using grayscale images. Do you thing the neural network received binary black/white pixels or the gray ones?

I do not know. 

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAMgOLLwkdPEWjteO1%2BJ0M0TVvfcfB9hvEX-3uQUvo_8dnr%2B2kw%40mail.gmail.com.

[Lorenzo Bolzani]

Yes, old OCR solutions use binarized content but I see this as a legacy limitation. It was probably done to speed up the processing and also, I suppose, because the algorithms used would not benefit from the extra gray details anyway. Old ocr tech was also print oriented so the text was already near binary.

With a neural network there is no extra time cost in processing grayscale or binary text, they are just float values in both cases. Binarization throws away a lot of data, especially with noisy images, complex backgrounds, etc. (ID documents, smartphone pictures, etc.).

Binarization may improve OCR performance but I doubt: a CNN should be easily able to learn to binarize the image itself if this can improve the results.

The only advantage I see for binarization is that syntetic training data is binary, so you try to match the real input data to the one used for training. Of course you could corrupt this data to make it grayscale.

But I would expect a full grayscale training and prediction to give slightly better results, especially for complex cases.

I fine tuned my models using grayscale data (real world crops, not synthetic) and, if possible, I'd like to try to disable the binarization step to see if I get an improvement. Maybe there are some parameters controlling this step.

Thanks

Lorenzo

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAJbzG8zBcMoYnt1WPjJffLVoRXoFZcBDu5QgnQ3gOyJTobi0aw%40mail.gmail.com.

[Lars Fricke]

Hi,

last week i did a lot of experimentation with binarization in my preprocessing. If you feed Tesseract with unbinarized data, it uses Otsu Thresholding. But even though documentation states that it uses adaptive (local) Otsu, i cannot find that in the code - but i am not an expert with c. And if you do _global_Otsu with "difficult because of low, varying contrast" input outside of Tesseract and feed the result to Tesseract, the output matches expectation. So i guess, tess4 defaults to global Otsu what really is a bad idea given the fact, that local binarizations should give better results in almost every case. 

So i used Fiji/ImageJ to check other binarization Methods. The one with best results for OCR with my data was the Phansalkar-Method. I advise you to test different binarizations with your data, then feed it into Tess4 and then compare results. Fiji-plugins are quite simple to modify, you can even test your own binarization method - i tried some variation of Phansalkar modifying the k variable depending on the local-mean to check if i was "on a line of text" - so far with pretty good results in very problematic regions with low contrast between text and background.

P.S: Despecle ans Skew-Correction filters might also help a lot, just play with them on a "typical" set of your data.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CAMgOLLwj%2B-xKQttOog-DCfii6TJp%2BqdWij8FpoeUpBgtJZH6Ww%40mail.gmail.com.