Letter Classification

[rland93]

Once we have localized a target, we need to figure out the letter. Remember, from #3 we will have a black and white (binary) array for the input.

For this, we have a couple ideas.

Letter detection is very well solved. One idea is to write a simple CNN with just a few layers and train it on our own data (there are plenty of binary letter datasets out there).

Another idea is to use a pre-trained network from somewhere. I'm not sure if networks trained on MNIST will work here, but there are a bajillion of those out there.

Another idea is to use a library like https://github.com/tesseract-ocr/tesseract to obtain letters.

In any case, there are many ways of doing this. Just keep in mind, we want this to be computationally very light, so work with tiny images (50x50 or less) with either binary or greyscale color.