Frequently grabbing images of that size and then running image detection on them in real-time seems like a very, very heavy task to do, cpu-wise. It would require a third-party library like Asprise or Tessaract and probably more than a little experience in programming.
Streaming programs or the plugins that generate the overlay don’t interact with the image data in the way that you would need, as far as I am aware.
There is probably a better way to accomplish what you are trying to do than to grab the numbers AFTER they are being rendered on the screen. What exactly are you trying to do?