Use convolutional neural network (CNN) instead of feedforward network
generic-github-user opened this issue · comments
Yeah this will be huge improvement.
This was the plan originally, I just wanted to make sure it was working with with a simpler network before making it more complex. It's not a complex problem anyway - pretty good results can be achieved by just comparing the number of white pixels to the number of black pixels. It could be made more interesting by making the sizes/colors of the circles variable/random.
just comparing the number of white pixels to the number of black pixels.
Yeah but wont help with overlaps
That's the point of adding mixed sizes/colors for the circles, so the model has to look at the shape of the circles in depth.