Tesseract 4 for old languages

[Guillaume Desforges]

Hi

I want to try using Tesseract 4 for old manuscript languages ("The Song of Roland" and such).

I have looked at https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00 but the steps are very unclear.

I have an image and a text file with the line content for each line of manuscript text. The doc says what to do, but not how.

I first thought I'd need img/box files pairs, but it seems it was for Tesseract 3 and is now irrelevant...

So I guess my starting point is here : https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00#tutorial-guide-to-lstmtraining

There is no tool to create the lstm-recoder directly. Instead there is a new tool, combine_lang_model which takes as input an input_unicharset and script_dir(script_dir points to the langdata directory) and optional word list files. It creates the lstm-recoder from the input_unicharset and creates all the dawgs, if wordlists are provided, putting everything together into a traineddata file.

I don't really get this part. How do I make  input_unicharset ? What is langdata?

Thanks

Guillaume Desforges

[ShreeDevi Kumar]

Please try tesseract 4.0.0beta.1  with languages such as

enm (English, Middle (1100-1500))

and

Fraktur  script 

Also, look at the following project from a few years back

http://emop.tamu.edu/outcomes/Franken-Plus

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

--
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-ocr+unsubscribe@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/fe1d68a2-76ce-4005-98ea-672710365517%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[ShreeDevi Kumar]

>I have an image and a text file with the line content for each line of manuscript text. The doc says what to do, but not how.

>I first thought I'd need img/box files pairs, but it seems it was for Tesseract 3 and is now irrelevant...

Tesseract4.0.0beta.1 does not officially support LSTM training from box/tif pairs. 

It uses box/tif pairs generated using the synthetic training data generation pipeline using a training_text and set of fonts, for making the lstmf files that are used by lstmtraining.

langdata refers to the langdata repository under tesseract-ocr github repo. The files in it have not been updated for 4.0.0

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

[jbc...@hotmail.com]

Same question here. I see that the documentation on training Tesseract 4 makes some reference to manuscripts:

    As with base Tesseract, there is a choice between rendering synthetic training data from fonts, or labeling some pre-existing images (like ancient manuscripts for example).

So, if I understand correctly, there is no support yet for training with labelled pre-existing images ? The concept of font does not makes sense with manuscripts, and what we can use in this case is just pairs of images and text (transcription).

Best,

Jean-Baptiste Camps

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

[ShreeDevi Kumar]

Please see the project https://github.com/OCR-D/ocrd-train

It has support for training tesseract if you provide line images and matching ground truth text. 

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/200db744-d010-4555-a4b7-86c64ba0b9bf%40googlegroups.com.

[ShreeDevi Kumar]

Please also see http://doc-creator.labri.fr/

which makes it easy to create synthetic data similar to manuscript pages.