2

Is it possible to increase the accuracy of specific characters in an existing traineddata model?

For example:

  • The number 3 will often be detected as 5.
  • The number 5 will often be detected as 8.
  • The W will often be detected as V.

Makes it sense to fine-tuning the eng.traineddata only with these characters, to increase the accuracy of these specific characters?

I tried to train the eng.traineddata model with line based image data and the accuracy for the characters in the example increased significant. But with the new model, letters that were well recognized in the previous model are recognized incorrectly. (Overfitting?)

Sean Stayns
  • 4,082
  • 5
  • 25
  • 35
  • Does it fail on specific fonts or in general? You could just train it on more fonts and more data (all characters) and see if that improves results – K41F4r Jun 08 '20 at 16:44

0 Answers0