CODEWITHSUNDEEP
machine-learningneural-networkimage-recognitionfeed-forward
Greg Peckory
Greg Peckory

Reputation: 8068

Neural Network Character Recognition

Suppose I'm trying to create a Neural Network to recognize characters on a simple 5x5 grid of pixels. I have only 6 possible characters (symbols) - X,+,/,\,|

At the moment I have a Feedforward Neural Network - with 25 input nodes, 6 hidden nodes and a single output node (between 0 and 1 - sigmoid).

The output corresponds to a symbol. Such as 'X' = 0.125, '+' = 0.275, '/' = 0.425 etc.

Whatever the output of the network (on testing) is, corresponds to whatever character is closest numerically. i.e - 0.13 = 'X'

On Input, 0.1 means the pixel is not shaded at all, 0.9 means fully shaded.

After training the network on the 6 symbols I test it by adding some noise.

Unfortunately, if I add a tiny bit of noise to '/', the network thinks it's '\'.

I thought maybe the ordering of the 6 symbols (i.,e - what numeric representation they correspond to) might make a difference.

Maybe the number of hidden nodes is causing this problem.

Maybe my general concept of mapping characters to numbers is causing the problem.

Any help would be hugely appreciated to make the network more accurate.

Upvotes: 1

Views: 205

Answers (1)

greeness
greeness

Reputation: 16114

The output encoding is the biggest problem. You should better use a one-hot encoding for the output so that you have six output nodes.

For example, 

-   1 0 0 0 0 0
X   0 1 0 0 0 0
+   0 0 1 0 0 0
/   0 0 0 1 0 0
\   0 0 0 0 1 0
|   0 0 0 0 0 1

This is much easier for the neural network to learn. At prediction time, pick the node that has the highest value as your prediction. For example, if you have below output values at each output node:

-   0.01
X   0.5
+   0.2
/   0.1
\   0.2
|   0.1

Predict the character as "X".

Upvotes: 3

Related Questions