how to improve prediction for heavily rotated image and special character

[ramdhan1989]

Hi,
I used small_satrn for these images and the model seems failed to recognize special character and heavily rotated image.

normal position with "-" symbol. it was successful to recognize the number only ((predict string "109")

heavily rotated image. it was not successful to recognize number (predict string "8")

it was successful to recognize normal position (predict string "185")

it was successful to recognize slightly rotated position (predict string "300")

how can I improve the model? do I need to retrain the model using heavily rotated image? why the small_satrn cannot predict special character even in normal position image while the config has included that character inside the config file?

thank you