0 is always recognized as 8

[nipunasudha]

I'm using the traineddata file from this project to recognize seven segment digits. Even for a very clear & high-resolution sample, 0 is recognized as 8. I searched through some stackoverflow, issue seems to be common. Take a look. http://stackoverflow.com/questions/30479002/digital-numbers-on-tesseract-ocr

[arturaugusto]

Can you send your sample, so I can investigate?

[nipunasudha]

These pre-processed images are actually photoshopped. But in the real experiment I used openCV to process. (I currently don't have them with me). Can you please improve the trained data not to limit to the font you used? I can help you find different types of seven segment samples. There are many looking for a good seven-segment OCR, yours is the best we have.

[arturaugusto]

You can have a idea of what the preprocessed image should looks like, on the trainning source image: [https://github.com/arturaugusto/display_ocr/blob/master/training_source/eng.letsgodigital.exp0.tif]() Note that the segments are expected to be connected. You can achieve this by using erode with opencv. I belive that tesseract don't works good with fonts where segments that are not connected.

[nipunasudha]

Sorry but erode isn't working sir. The exact answer & reply for your suggestion, snapshot from stack-overflow http://stackoverflow.com/a/30604944 (this answer is for your traineddata)

[arturaugusto]

Just did some tests with your image printed:

I can see that 0 is kind of problematic. Don't know if other trained data solved the '0' issue without break other characters, but I will take a look. Maybe with some changes on trained data image source.

[nipunasudha]

Yes please! and also can you place the decimal point after random digits in the training image? that helps recognizing the decimal point. Also a MINIMAL VERSION (with 0..9 and decimal point) of the trained data would be awesome. (tesseract can be configured to achieve that though) Your work is extremely helpful, thank you a lot.

[nipunasudha]

Just a reminder, sir could you take a look to solve this issue? It would be a great help for me, and many others.

[arturaugusto]

Sorry @nipunasudha. I'm quite busy right now. If you wanna see this working very soon, I recommend you to follow tesseract docs and try changing the box files to see if the issue is fixed. Right now, I can't say to you when I will be able to do this. Would be nice to also create some tests.

[arturaugusto]

looking into this today.

[nipunasudha]

Thank you so much!

[dpanic]

any updates?

[arturaugusto]

Try using this trained data: https://github.com/Shreeshrii/tessdata_ssd/blob/master/7seg.traineddata

[dpanic]

I already did.

I dont want to use other people trainings, which are done by legacy Tesseract 3.X

On Thu, 14 Oct 2021 at 13:20 Artur @.***> wrote:

Try using this trained data: https://github.com/Shreeshrii/tessdata_ssd/blob/master/7seg.traineddata

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/arturaugusto/display_ocr/issues/6#issuecomment-943262504, or unsubscribe https://github.com/notifications/unsubscribe-auth/AACUTPNKA6PDODU2TN2BF43UG24IRANCNFSM4DDQEXAA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

--

-- Dušan Panić

Senior Security Engineer

Email: @.** Mobile: +381 60 374 58 11 GitHub: https://github.com/dpanic Twitter: https://twitter.com/dusan_panic LinkedIn: *https://www.linkedin.com/in/du%C5%A1an-pani%C4%87-5933731b2/

"Tell me and I forget. Teach me and I remember. Involve me and I learn." ~ Benjamin Franklin

[arturaugusto]

Also, I'm working on this: https://arturaugusto.github.io/7seg-ocr/ Maybe a alternative that dont need tesseract

[dpanic]

Thank you Artur. I will check it out. However I must figure out how Apple does OCR on iOS, they do it pretty good.

On Sat, 23 Oct 2021 at 16:08 Artur @.***> wrote:

Also, I'm working on this: https://arturaugusto.github.io/7seg-ocr/ Maybe a alternative that dont need tesseract

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/arturaugusto/display_ocr/issues/6#issuecomment-950158400, or unsubscribe https://github.com/notifications/unsubscribe-auth/AACUTPJITIIIV72VTJNGF7TUIK6WPANCNFSM4DDQEXAA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

--

-- Dušan Panić

Senior Security Engineer

Email: @.** Mobile: +381 60 374 58 11 GitHub: https://github.com/dpanic Twitter: https://twitter.com/dusan_panic LinkedIn: *https://www.linkedin.com/in/du%C5%A1an-pani%C4%87-5933731b2/

"Tell me and I forget. Teach me and I remember. Involve me and I learn." ~ Benjamin Franklin

[arturaugusto]

7 segments is harder to OCR using trained models if you need something near 100 % accuracy, that why I'm working on analytical solutions. There is also this one (not open source): https://ocr.ipt.br/?font=7seg. It uses tesseract ocr compiled to wasm.

[dpanic]

@arturaugusto Thanks. I already managed to get with your letsgodigital.traineddata to archive 94% accuracy. But problem is it is too slow (long story)...I perform multiple OCR's on multiple images, than I sumarize results.

I am building my own training...

Btw your https://github.com/arturaugusto/display_ocr doesn't contain all data with which you built ~10mb traineddata :-)

I am using your scripts, and Tesseract 3. I end up with 140kb file, which has 70% accuracy...Around 1000 pictures (combined) into multiple Tiffs for training.