Improved letter recognition results
95% success rate with 0% misidentified letters
Less is more
I'd been trying to recognize the same set of letters (including the ugly ones from the previous entry) with increasingly complicated neural networks. It worked at first, that produced the 75% success rate that I posted in the edit of my last entry, but then I was getting numbers around 62%; totally unacceptable. A 3 layered neural network gave me around 61%.
So I went back to an architecture with one hidden layer, and instead only changed the number of nodes in it to 100, up from 43 in my original test. Again, I did not remove the "bad" letters from the images.
And I got a 95% success rate with 0 erroneously identified letters :D
The per-letter error rates are:
So the results for the B are terrible, bad for the H and acceptable or good for everything else.
All of these results are based on a 516 letters test set.
Less is more
I'd been trying to recognize the same set of letters (including the ugly ones from the previous entry) with increasingly complicated neural networks. It worked at first, that produced the 75% success rate that I posted in the edit of my last entry, but then I was getting numbers around 62%; totally unacceptable. A 3 layered neural network gave me around 61%.
So I went back to an architecture with one hidden layer, and instead only changed the number of nodes in it to 100, up from 43 in my original test. Again, I did not remove the "bad" letters from the images.
And I got a 95% success rate with 0 erroneously identified letters :D
The per-letter error rates are:
- B - 54%
- D - 11%
- G - 5%
- H - 22%
- M - 3%
- R - 18%
So the results for the B are terrible, bad for the H and acceptable or good for everything else.
All of these results are based on a 516 letters test set.