GOCR to Neural Nets Pt 2
As per usual these posts don’t go as smoothly as I would like. The idea was to use FANN for PHP to make a simple neural network that would work easily. Hahahaha. You might think I make this stuff up as I go along. Oh wait. I do.
Anyway FANN requires PEAR to be installed and I figured that it’d be much simpler than installing PEAR modules to find something that was completely PHP to do the job. I did that. However.
1. It’s slow.
Ok so it’s slow. We can live with that right? I mean we’ve got some time.
2. It’s slow.
Ok it really is slow, and I’m getting impatient.
3. It’s painfully slow.
PHP sucks for some things. I do like C’s direct memory access. On the plus side this little neural net class is so simple and easy to understand.
Anyway I went ahead and I fitted all my modules together around this PHP neural net and I walked away whilst my computer attempted to learn the alphabet at a ridiculously slow pace. 900 captchas later I give it up.
The neural net does sort of work now for quite a few characters. Although I did notice a character failed to segment properly which set the learning back a bit. A lot of the characters it fails on are things like producing P’s for R’s or vice versa. So you can understand where the problem lies.
I won’t post code because there’s a fair bit spread over a few modules. nnbreak.php is the main module and must be run from the command line. Like this:
php nnbreak.php captcha.jpg answer.txt train=1/0
1 means train the network using the answer stored in answer.txt. 0 means guess what the captcha is in captcha.jpg, which will ignore answer.txt but answer.txt must still be included (although it can be blank or not exist etc).
So analysis… Here’s how it works roughly:
Allocate memory for the neural network, and load in previous neuron weight values
Extract the letters as said in that post about extracting letters.
Convert the letters into a 10×10 matrix of averaged values, maximum being 1.0
Loop through each letter and send the matrix of values to the input neurons
Check the output neurons
Possibly teach the network the right answer using back propogation.
We can tweak a lot of things such as the number of neurons in each layer and the size of the matrix. I did mess with the default backpropogation teaching speed because it didn’t seem to be learning fast enough to me 
. I’m guessing that has some drawbacks to it but I’m not sure exactly what they are. In this line the 0.5 is the learning speed which has been moved up from the default 0.1:
$nn = new nn(3, $layer_structure,1,0.5,0.9);
Now if anyone happens to train the network to recognise all of the letters properly send me the file so I can claim it as my own and pretend I did it all perfectly 
. J/k. Seriously though I do think it should eventually learn all the letters properly it’s just taking a long time.
In conclusion. I don’t trust neural networks they do too much stuff that I don’t know about. I’m betting that robot from the Terminator was probably built from neural nets.
The code to guess a captcha using neural nets - Already populated with some weights so it sort of works
Sunday, March 30th, 2008
