Digits reading optimalisation.

[Benek]

Hello! I'm trying to read some digits and I thought it was a rather simple task yet still I can't receive satisfying results. So my first question is: is it possible to get 100% accuracy when reading some standardized input? Or there will be always some errors when reading?

Here are some sample inputs that I wanted to read:
The digits that are being misread are:
on the photo t3
5.1 is read as 9.1
on the photo t4
: 10.2 is read as 610.2

I'm using:

tesseract/4.1.1

config :

oem: 3,

    psm: 11,

    tessedit_char_whitelist: "0123456789.",

    load_system_dawg: false,

    load_freq_dawg: false,

The images have 2700x2100 resolution.

The 999 on the left are markers that I added to be able to recognize which line belongs to which output text and they are always read correctly.

I tried experimenting with some different image preprocessing techniques like blur, median, changing the size of the image etc.

Do you have any other tips that could lead to better reading accuracy?
Thanks in advance for any help!

[Владимир Калачихин]

Heh. It's an old issue.

For 100% accuracy, you must use a digit-only language model. But there is no such thing.

Besides trivial perceptron shows good results on digits recognition.

суббота, 30 января 2021 г. в 18:41:13 UTC+3, Benek:

[Benek]

I still need to read the dot in the correct place which makes it a bit harder. So you don't think it's a problem with the input data?

[Владимир Калачихин]

Digits included in language model with letters. And model most trained to phrase recognition, not separate digits. Mistakes on digits unavoidable.

суббота, 30 января 2021 г. в 19:12:39 UTC+3, Benek: