Description

• How this works: this takes in an image, reads and preprocesses one frame by isolating the box containing the digits and then isolating each digit. Then it feeds it into a forward NN to predict each digit and sums the values according to their weights(tens, ones, tenths)

Limitations

- Bounding box padding sometimes inaccurate with cutoff before the decimal place

• Possible improvements:
  • Most important:
    • Using a 4 point transform so that the numbers can be viewed from an angle
    • Model not very accurate
  • Making bounding box detection faster? Currently I’m looping over all pixels and that could be potentially really slow
  • Replacing the feed-forward net with a CNN

Images:

Testcases: WIP

• I wrote a simple test case with unittest that tests its capability to detect the right numbers - currently failing