Merged from my googlecode repo.
A Java Number Recognition Program based on Convolutional Neural Network
This test project is based on the document of Neural Network for Recognition of Handwritten Digits By Mike O'Neill (http://www.codeproject.com/KB/library/NeuralNetRecognition.aspx).
In the project, a trained weights set (60000_9.nn) is provided, which is trained from the THE MNIST DATABASE of handwritten digits(http://yann.lecun.com/exdb/mnist/index.html).
For the test set (10000 test case) from this website, it can get a error rate of 4.26%
The following screenshot shows the verify UI:
License
This is LGPL license so everyone can use it freely. The code can be changed if needed. I have to say that there may be bugs or unexpected behaviors (Exceptions, Errors) of the program. So use it on your own risk and test throughly.
Requirement
JRE 5.0 or later. I debug and test with JDK6, so I recommond JDK/JRE6.
The GUI part requires a Window System (Microsoft Windows, KDE, Gnome, etc). But the core part does not.
Usage
Unzip and execute run.bat (Require java.exe on PATH)