2
$\begingroup$

I have built a program that recognizes shapes and characters, using Neural Networks. Now, one of the main requirements of the task is that the program can recognize them regardless of rotation or mirroring. So far it has worked well. To do this I used Fourier Descriptors as input data to the NN.

I am now implementing a similar program but in a embedded platform with limited resources (memory, speed,etc). I have just started.

I could try and implement the same thing I did (obviously the training part will have to be aided by a PC-I suppose) but thinking about the limited resources and not wanting to do things more complicated than required, (and plus I read a book where they recommend almost none image preprocessing- just input the pixels into the NN) I have the following question(pardon if it is naive or obvious)

If I train a NN with the inputs being a character or shape image pixels, would it be able to recognize later a rotated version of it?

What if I train rotated versions of the characters as well?

Or maybe it is necessary for me to use Fourier descriptors after all? Any recommendation to build a NN that recognize rotated versions of characters or shapes?

$\endgroup$

1 Answer 1

1
$\begingroup$

If you, during the successful training if the NN model, use data augmentation in the form of rotations and mirroring then they will become invariant to the model.

Namely it will have to extract features which are invariant to those transformations.

This will result in a model which will be able to do what you need.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.