How do I only detect text of one size?

[Radu Stoicescu]

I have some scanned, machine typed, that have a lot of noise. I can reduce the noise, and I have done so. But there is some noise that is statistically indistinguishable from letters: as dark as the letters and as big as the letters, therefore I cannot just take it out.

I have tried to only train Tesseract on Courier New, and although the accuracy went down, which was expected because I did not use enough data, there were still letters detected in the noisy areas.

How can I keep Tesseract from detecting letters in noise? One simple rule would be to only detect characters of one size, since this is machine typed text.

[Zdenko Podobny]

Maybe it would be good to provide some examples of input.

Zdenko

pi 25. 9. 2020 o 7:57 Radu Stoicescu <radust...@gmail.com> napísal(a):

I have some scanned, machine typed, that have a lot of noise. I can reduce the noise, and I have done so. But there is some noise that is statistically indistinguishable from letters: as dark as the letters and as big as the letters, therefore I cannot just take it out.

I have tried to only train Tesseract on Courier New, and although the accuracy went down, which was expected because I did not use enough data, there were still letters detected in the noisy areas.

How can I keep Tesseract from detecting letters in noise? One simple rule would be to only detect characters of one size, since this is machine typed text.

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/24c0b6ae-e07b-443b-ba60-38470b852275n%40googlegroups.com.

[Radu Stoicescu]

The first image is OCR before any pre-processing, the second image is after pre-processing. As you can see there are a few problematic areas. I can understand that there is very little to be done where the line-like noise is confused with an underscore but the 2 areas where the 2 "e" and the 1 "e" are detected something could be done.

As I said, I tried to retrain the top using only "Courier new" but the noise was still detected as letters, that was surprising to me. I thought that the false positives are because of the large amount of different and strange characters Tesseract is trained on.