Which representation of .bmp files is effective for pyBrain network to teach it recognising simple images? I am not sure plain feeding of one-dimensional sequence of pixels is good enough.
Asked
Active
Viewed 71 times
1 Answers
0
Your assumption is correct. Just imagine what a 1 degree rotation would do.
The problem isn't in .bmp files. That is an utterly trivial detail. The real problem is in the set of transformations for which you may need to correct, and we simply have no idea. No name just a few: translations, magnifications, rotations, perspective, skew, tilt, blurring, depth of focus, ...
To take just the first, translations, you'd compensate by moving the center of gravity to the center of your image.
Of course, if you intend to distinguish d and p, you'd better be careful in eliminating rotation differences, but otherwise you can normalize the major axis to 0 degrees.
MSalters
- 173,980
- 10
- 155
- 350
-
Just a question: if you wanted to pick the "right" choice between `d` and `p` or `b` and you are recognizing letters from words, is that where Markov chains might come in to help you pick the most likely letter given the current or current and previous letters? – Engineero Jun 12 '14 at 15:07
-
I have pre-generated simple files with one digit in centre. Nothing complicated. – RomaValcer Jun 12 '14 at 18:44