How to optimize tesseract to maximum speed for single number (several digits) recognition

[Jan Pohanka]

Hello,

I'm making a simple device used to recognize numbers on pictures taken by a webcam. All is running on raspberry pi 3.

Everything is like following simple loop (in python for simplicity, but using C++ api it is the same), images are preprocessed to black and white

api = PyTessBaseAPI(psm=tesserocr.PSM.SINGLE_WORD)

for im in images:

    api.SetImage(im)

    api.SetSourceResolution(70)

    ot = api.GetUTF8Text()

api.End()

My problem is that api.GetUTF8Text() call is quite slow and more over it is getting slower and slower over time. Is there any options how to make recognition faster? I have tried to resize the image to around 50x10px. The times starts on around 300ms but then goes up to above 1s which is too slow for me. I tried both legacy and LSTM algorithms, but they are similar.

best regards

Jan

[Lorenzo Bolzani]

First double check if the Pi is not throttling due to overheating or lack of USB power. This may cause the slowdown.

Usually 30/50 px of text height is fine. IF the problem is tesseract, try to use the fast model (or "normal" if using best). I assume you are using the 4.x release.

Try tesseract -v to see if you are using all the available CPU optimizations.

Try to move the SetSourceResolution outside the loop and see if it changes something (MAYBE it may invalidate some caches or something).

The time you are referring to is one single api.GetUTF8Text() call, correct?

Lorenzo

--
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/a53b4b25-97e3-47dc-823a-cbb219225eed%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jan Pohanka]

Thanks for suggestions. You are right that I'm reffering to api.GetUTF8Text() call, it is my bottleneck.

I was not aware that there is some fast and best models in tesseract 4.0, I will give it a try. So far I used just lang=eng or osd.

For me it is suspicious that the calls are getting longer during the time. Or to be more precise, first 10-15 calls are up to 500ms and latter ones rise above 1s...

SetSourceResolution outside of the loop gives no change unfortunately.

BR

Jan

Dne úterý 29. ledna 2019 18:08:49 UTC+1 Lorenzo Blz napsal(a):

[Zdenko Podobny]

What is your tesseract version?

Zdenko

ut 29. 1. 2019 o 20:40 Jan Pohanka <xhpo...@gmail.com> napísal(a):

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/baa59c86-b002-4607-8dda-16835cd3ea73%40googlegroups.com.

[Jan Pohanka]

It is 4.0. I'm satisfied with recognition results, but I need to make it faster (at constant times below 1s)...

Dne středa 30. ledna 2019 7:48:23 UTC+1 zdenop napsal(a):

[Zdenko Podobny]

search issue tracker for "speed"...

Zdenko

st 30. 1. 2019 o 7:51 Jan Pohanka <xhpo...@gmail.com> napísal(a):

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/dedb7fd8-d61e-42bb-a492-34beaa8b1514%40googlegroups.com.

[Jan Pohanka]

I have already done that but haven't found anything interesting.

I tried to ask here if there are eg. any part of algorithms that can be disabled etc. The image is preprocessed, binarized and contain only 8 digits (and point). I was also a bit surprised that resizing image from 400px to 50px has given only subtle speed up.

I will try the fast model today (if I find how to switch it), maybe it will help.

here are my measured times

ocr time: 0.980876922607

ocr time: 0.435426950455

ocr time: 0.76907491684

ocr time: 0.836761951447

ocr time: 0.871710062027

ocr time: 0.803520917892

ocr time: 0.371052026749

ocr time: 0.732284069061

ocr time: 0.745162010193

ocr time: 0.836426019669

ocr time: 0.740739107132

ocr time: 0.379159927368

ocr time: 0.798940181732

ocr time: 0.3972260952

ocr time: 0.739762067795

ocr time: 0.7757999897

ocr time: 0.772871017456

ocr time: 0.435608863831

ocr time: 0.770547866821

ocr time: 0.870738983154

ocr time: 0.37126493454

ocr time: 0.837875127792

ocr time: 0.811723947525

ocr time: 0.865257024765

ocr time: 0.79048204422

ocr time: 0.435704946518

ocr time: 0.763910055161

ocr time: 0.391008853912

ocr time: 0.396636009216

ocr time: 0.38174700737

ocr time: 0.809095144272

ocr time: 0.773195028305

ocr time: 0.427488088608

ocr time: 0.403608083725

ocr time: 0.806233167648

ocr time: 0.948635101318

ocr time: 0.900885105133

ocr time: 0.829130887985

ocr time: 0.932774782181

ocr time: 1.09788799286

ocr time: 0.520708799362

ocr time: 0.448786973953

ocr time: 0.560626983643

ocr time: 0.993177175522

ocr time: 0.48442697525

ocr time: 1.1292309761

ocr time: 1.04695606232

ocr time: 0.8810338974

ocr time: 1.10285806656

ocr time: 1.05213713646

ocr time: 1.22593903542

ocr time: 1.04618191719

ocr time: 1.11645102501

ocr time: 1.05435395241

ocr time: 1.15162396431

ocr time: 0.547721862793

ocr time: 0.607867956161

ocr time: 1.14074802399

ocr time: 1.1790971756

ocr time: 1.18815803528

ocr time: 0.58503985405

ocr time: 1.10898280144

ocr time: 1.22723913193

ocr time: 1.2178709507

ocr time: 1.28540086746

ocr time: 1.28237104416

ocr time: 1.56176805496

ocr time: 1.2859480381

ocr time: 1.2599170208

ocr time: 1.42588591576

ocr time: 1.51333785057

ocr time: 1.34276986122

ocr time: 1.34283900261

ocr time: 1.39351201057

ocr time: 1.61450195312

ocr time: 1.44723105431

ocr time: 1.63176107407

ocr time: 0.82429599762

ocr time: 1.08239603043

ocr time: 0.755813121796

ocr time: 1.63984704018

ocr time: 1.84553313255

ocr time: 0.958009958267

ocr time: 1.52479290962

ocr time: 0.919597864151

thanks

Jan

Dne středa 30. ledna 2019 7:57:07 UTC+1 zdenop napsal(a):

[Lorenzo Bolzani]

Did you check this?

https://www.raspberrypi.org/forums/viewtopic.php?f=63&t=147781&start=50#p972790

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/30831c1e-4bd9-4eae-8118-ad9244949b80%40googlegroups.com.

[Jan Pohanka]

You were right, I just found that my RPi is throttling. It explains the slowing down. Now I'm checking if heatsink could help.

So I expect that there is nothing to tune up in my loop. I will check if I can try some smaller model.

best regards

Jan

Dne středa 30. ledna 2019 11:15:19 UTC+1 Lorenzo Blz napsal(a):