Does the psm value used to generate lstmf files influences the training?

[Lorenzo Bolzani]

Hi,

I keep having problems with duplicated letters with custom fine-tuned models.

For example an M becomes MH.

I'm using ocrd-train with actual crops and I noticed that the lstmf files are generated with psm 6.

At runtime I use psm 7. Do you think this may make a difference? From a quick test it does not seem the case.

The problem gets worse if I use psm 13 for recognition this is why I'm wondering if there is a relation.

Is there something else that I'm doing wrong that might lead to this problem? Or something I can improve?

I have only one font (ocr-b) with fixed height (44px plus 2px white margin).

According to this post the sweet spot seems to be closer to 30px (for most fonts)

https://groups.google.com/forum/?#!msg/tesseract-ocr/Wdh_JJwnw94/cHjYD3cDEQAJ

Thanks, bye

Lorenzo