incorrect image classification using NN

hi all i have implemented a neural network to recognize printed character (0-9 and L R ), I got correct classification when testing in offline mode i.e. on previously captured images but when i connect the camera and test the NN in the online mode i got completely incorrect classification i used wavelets as features extractor with resolution 4 and i noticed some differences in features extracted during online capturing and the straining one , can anyone help me out figure the error or advice me with something

[Image Analyst]

How could we do that, not knowing anything at all about your images, code, or algorithm? See the paper clip icon? Use it to attach any files you think might be useful to someone who may want to help you.

[Sarah Mahmood]

ok your right .... I attached the m file code and a sample image label

[Sarah Mahmood]

no one can help in this !

[Image Analyst]

Well, there aren't that many NN experts around. Perhaps you'd be willing to use more traditional methods, like here. Anyway, OCR questions never get a lot of help in this forum - they're just too complicated and involved for us to answer. Most people want a turnkey OCR program just handed over to them, and we just can't do that. We can help on small snippets of code only.

[Sarah Mahmood]

yes thank you, but I don't want a code I implemented my own but I want an advice if there is something I'm missing because when this program tested on the trained images it's work perfectly but in real time acquisition of image I got wrong recognition that's all

Answer 1

[Greg Heath]

Test the program offline on non-training images before testing it online.

I suspect you have over-trained an over-fit net so that it essentially memorized the training images but is not able to generalize to non-training images.

How many input/target examples do you have? N = ?

What is the dimensionality of your input feature vectors? I = ?

[ I N ] = size(input)

What is the dimensionality of your output/target classification vectors? O = ?

[ O N ] = size(target)

Since you have c = 12 categories, O = c = 12 with target matrix columns equal to c-dimensional unit vectors with the row index of the 1 indicating the true class index of the corresponding input vector.

What are the sizes of the train/val/test sets? Ntrn/Nval/Ntst = ?

How many hidden-layer nodes? H = ?

What is the ratio of training equations Ntrneq = Ntrn*O to unknown weights Nw = (I+1)*H+(H+1)*O ?

[Sarah Mahmood]

thank you, I tested it offline with non trained data and I also got wrong classification the answers to your questions are 1- I have 20 imaged labels which classified to 20 textual labels 2- the dimension of feature vector is 4x5 as each label consisted form 4 patterns and I have extracted 5 features for each. 3- size of input is 80x5 (for the whole labels 20x4=80) and size of output is 1x4 which corresponds to the 4 patterns of the label 4- I used two hidden layer of 20 and 30 neurons for the c=12 categories and the sizes of train/val/test sets? Ntrn/Nval/Ntst = ? I didn't get them I'm sorry but I'm little new on this field if you kindly explain more on them