Question

I need to find some method to recognize images with a single digital number in it (black on white).

What I have read so far, I would use a neural network. Is a SOM (self organizing map) the correct network type?

Is there any other good way to to recognize digital (I have found a lot of stuff for handwritten characters, but not on digital numbers) numbers that I haven’t thought of yet?

Answer 1

I think the type of neural network you´re talking about is a multiclass classifier, right? So you have some binary input units(maybe colors from white to black) and you expect a binary output, for example that your image shows a three. I´m not quite shure about that but I think one of the best methods to do that is a deep belief network. They were tested with the MNIST data set(should be aviable somewhere in the web if you search for it..) and they worked pretty well. The training process is a little complicated, you first train the network unsupervised layer by layer and than apply the backpropagation algorithm(cross entropy error in that case I think). Ok, I´m currently working on a program for that task(in C sharp so it shouldn´t be to hard for you to understand). If you want the code contact me here: http://compositedevtec.tk/blog/contact/, or search youtube for “Hinton”(invented these models). There´s a video about him explaining his technique.
Hope this helped you a little bit