Page 1 of 1

OCR neuralnet make me prb!!!

PostPosted: Wed May 11, 2011 9:43 pm
by vattic
hi , andrew i need some help

I'm working on "hand writting recognition with Neural net" my algorithme at an advancead stage,

but i'm bloked in stage of Train that make to me a probleme i do the Train with BackPropagate algorithme on all 26 letters English language

1- I do the Train of 'Letter A' & adjust Weights and save this Weights_1 adjusted ;

2- load Weights_1 and di the Train of 'Letter B' and save new Weights_1 adjusted ect....
in stage of recignise i do local matching betwen the form and my data base(has desire Letters);

but in the recognize of letters the result be false coz the net save the last adjusted Weights (this is what i'm thinking)

so my net can only recognise one letter when my net do the matching.(when i do the train of 1 letter the result be good)

if you can help me I will be grateful to you.

i dont know where is the probleme?????

Re: OCR neuralnet make me prb!!!

PostPosted: Thu May 12, 2011 8:36 am
by andrew.kirillov

If you train your network providing letters one by one, then it will never work. You need to provide all characters you have to the network. So let’s suppose we speak about training epochs (you will need a lot of them to train network). On each training epoch you provide all 26 characters to network, calculate error for each and average. If error is too big, then start another training epoch, where you provide all characters again.

The above idea is the base of training network with error back propagation algorithms. So I would recommend getting back to theory.

P.S. Don’t need to write PMs with questions, which are supposed to be discussed on the forum. It does not really help.