minimum requirements for tesseract

[Yura Komlyk]

Hello!
I'm interested in system with minimum computational possibilities (the most 'slowest' CPU and minimum RAM) on which tesseract could still work.
The goal is to use it in minimized pc-platform device.
Maybe someone know another way to make minimized device with OCR capability. BTW only digits recognition is enough here.

Thanks.

[Leonardo Gabrielli]

I'd like to bump this topic.

I am new to OCR and I find weird enough that there's no hardware for OCR nowadays (dedicated ICs), no libraries for embedded architectures and the only option is to use ARM platforms with linux and tesseract. 

BTW running a test on a digit-only image on a raspberry Pi (armv6) with debian, I needed 34Mb of virtual mem, 1.5s of 99% CPU load. Embedding it into low capabilities cores can be surely hard.

[Nick White]

On Wed, Mar 12, 2014 at 09:20:02AM -0700, Leonardo Gabrielli wrote:
> I'd like to bump this topic.
> I am new to OCR and I find weird enough that there's no hardware for OCR
> nowadays (dedicated ICs), no libraries for embedded architectures and the only
> option is to use ARM platforms with linux and tesseract.

There is an EMBEDDED flag, but it isn't used much at the moment. The
GRAPHICS_DISABLED flag may also trim some of the memory load, if
you're lucky. Pass --enable-embedded=yes and --enable-graphics=no to
./configure to try them out. Are you stripping the binary? Also I
don't know if the raspberry pi has an OpenCL capable graphics card,
but if so you could enable opencl support, which would reduce the
CPU load.

Nick

[Robert Komar]

On Wed, 12 Mar 2014, Leonardo Gabrielli wrote:

> I'd like to bump this topic.I am new to OCR and I find

> weird enough that there's no hardware for OCR nowadays
> (dedicated ICs), no libraries for embedded architectures
> and the only option is to use ARM platforms with linux and
> tesseract.

OCR is still a black art. The algorithms are still under
research, changing constantly, and still not good enough
to cast into silicon. It's too early yet.

Cheers,
Rob Komar

[Leonardo Gabrielli]

Yep, I understand. Still I'd be curious to know how to achieve something as small and low power as this http://www.fastforward.ag/eng/index_eng.html 

Or similarly http://www.xemtec.ch/15-0-AMR-Retrofit.html employs an ASIC with image sensing and pre-processing but the recognition is left to a later processing (even on a 8-bit MCU!?)

Cheers,

LG