1

I'm working on an homemade ocr and I need some precisions that I didn't found on the web. I have in input an image of 50 x 50 pixels that I parse by pixels, if the pixel's color is more white I put 0 in a list and if it's more black I put one.

After that I give my list as the input for the neural network. So the size of my list is 50*50 = 2500 (is it too much ??). I have 3 layers : input, hidden, output.

I give as output during the training the ascii value of the caracter I send as input.

I use a sigmoid function for activation and the backpropagation for update the weights.

Currently my neural network doesn't work and my output, during training, is always increasing maybe because of the 2500 inputs which is to much ?

Every advices are welcome.

Thanks.

EDIT : Maybe resize the image by 16x16px in order to have 256 neurone in input layer ?

taspai
  • 71
  • 6
  • Possible duplicate of [training feedforward neural network for OCR](http://stackoverflow.com/questions/9684204/training-feedforward-neural-network-for-ocr) – Vaune Oct 30 '15 at 14:44
  • You can also look into multiple passes over the input data down sampled – Vaune Oct 30 '15 at 14:44

0 Answers0