OCRopus / ocropy updates

[Tom]

I joined Google this year. Google permits me to spend time on the OCRopus project and contribute. As part of this, I moved the project to Github, because it's easier to maintain there.

I just pushed out a new update of ocropy. This includes mainly faster/smaller saving of models, as well as a C++ implementation of the LSTM network. The C++ LSTM implementation is a pretty straightforward port of the Python version and runs much faster. The C++ classes have been wrapped as Python classes and are callable from Python. There are two new top-level drivers, ocropus-ltrain and ocropus-lpred, for the C++ implementation. The C++ implementation appears to be numerically close to the Python implementation and yield good recognizers when trained, but it requires more testing.

As before, this is research-level software with minimal documentation (do look at the iPython Notebooks, the .ipynb files, since they contain significant information). Feel free to contribute patches, documentation, etc. using the usual Github mechanisms of merge requests. I'll try to incorporate them as time permits.

Tom

[Chris Pietsch]

Cool, Tom!

I guess this is your GitHub repository: https://github.com/tmbdev/ocropy

Cheers!

[81+ yrsold]

Tom,
I am really happy - you have resumed ocropus project again. Trust this time I hope Ocropus Project will support for Indic lang(Indian languages) this time.
With warmest regards,
sriranga(81+yrs)

[Tom]

With the new recognizer, it should be pretty easy to train. We've trained it for other scripts purely from generated data and gotten pretty good results.

I'll try to create some more documentation and some simpler training scripts.

Tom

[Shibamouli Lahiri]

Hi Tom,

Thanks much for the update. I'm new to Ocropus, and I had a question on running rtrain.

Do you know (or have an estimate of) how many lines of text does the program take (to train) before it starts giving reasonable results? I'm wondering because since it's neural network based, I'd hazard a guess that it'd take more than a few thousand lines?

More details:  I'm working on gathering labeled data for Bengali (Bangla) OCR, and needed an estimate of lines that I'll need to transcribe as a starter.

Regards,
Shibamouli

[Tom]

Actually, it takes surprisingly little data: after a few thousand lines of text, you already get pretty readable results for Latin text.  

You can train on simulated data as well with good results: a tool for generating training data artificially is included (but probably requires a bit of adaptation for other scripts).

Tom

[Shibamouli Lahiri]

Thank you very much Tom for the information. I'll take a look into the simulation tool.

Happy New Year,

Shibamouli

[Everest]

Hello Tom,

      Does the C++LSTM implements ocropus-gpseg as well? I kind of need that function in my work currently. Thank you very much!

[Laurent Denoue]

Tom,

I noticed the en-default.pynn.gz file is huge, 80Mb, 350Mb uncompressed.

I thought neural networks were just small sets of weights. Why are the models so big?

Laurent