Where to start Handwritten Recognition using Neural Network?

Question

I\'ve been trying to learn about Neural Networks for a while now, and I can understand some basic tutorials online. Now i want to develop online handwritten recognition usin

Answer 1

Neural nets need a lot of soak time. The concepts are easy enough, but they can be overwhelming to the beginner.

Take a look at what Jochen Fröhlich has done with neural networks in Java. It sounds like an ideal starting point for a Java programmer like yourself.

Answer 2

Bear in mind that if your aim is to actually recognize these characters, your performance will stand and fall based on the quality and selection of the input features.

It is absolutely critical to choose the right features, and to preprocess (i.e. get rid of noise features, extraneous data, duplicate or strongly correlated features) as much as you can. In my experience, you'll get much better performance from the most boring and plain nearest neighbour implementation with good features than from a cutting edge algorithm with less well selected features.

For you, that means delaying reading the neural net literature for now (just take some off the shelf black-box implementation first) and reading up on what kind of image processing etc. real systems use. If your data can include pressure and speed info, all the better. Something like an LDA heatmap can be illustrative to initially see which features matter, and which don't.

For basic classification, there are tons of decent algorithms. Most work just fine and will work just fine for you. The difficult part isn't in picking or tweaking the algorithm, it's in avoiding the garbage-in-garbage-out scenario.

Answer 3

Heaton research will help you a lot

http://www.heatonresearch.com/articles/7/page3.html - visit this for tutorial with sample code snippets

Answer 4

Neural network [if i am not mistaken] would work on interpreting patterns, You provide inputs to your program and the program searches that pattern in a set of stored patterns and based on the match provides a possible match.

In your case, the dimensions for the patterns could be data such as speed + direction or only direction, etc

I made a small simulation of a neural bot that chats on my site in a similar fashion.

The more patterns the program 'Learns', the more accurate answers it provides.

Answer 5

Start simple with character recognition on the Unipen database.

You will need to extract pertinent features out of raw trajectory data in order to form what's commonly called a "feature vector". For instance you could resample the data using an interpolation scheme to end up with n tuples, each tuple containing information such as:

• position
• orientation
• velocity
• acceleration
• curvature
• etc

Once you have a fixed size feature vector, you use it as the input to your neural network. Try MLP networks for a start.

You will have to experiment in order to decide which features are best.

If you need to get started on extracting features from Ink data, have a look at HP's Lipi Toolkit (note that their recognizers don't use neural networks though).

You can also have a look at this 15 Steps to Implement a Neural Network tutorial.