Recognition tuning on text from a game

[Juan Pablo Morales]

Hi all,

I am trying to build some graphs from replays of starcraft. This is the kind of image that I use and I have croping functions that get each zone of interest (minerals, gas, workers, etc):

I do the following steps:

- I invert with img = 255 - img (img being a numpy array)

- I run image_to_string(stat_img, config='--psm 6 --oem 1')

Results without inversion are really bad, whilst when using inversion results get quite better. 

Still I get many times letters instead of numbers (e.g 8B instead of 80 or ee instead of 0), and I feel this could be improved since the data source is so standard.

What do you think I could add to this pipeline to improve results?

Thanks!