The input data of the neural network, the coordinates of the x / y coordinate plane, correlate with the handwriting

Question

I am very curious to create a handwriting recognition application in a web browser. Users draw a letter, ajax sends data to the server, the neural network finds the closest match and returns the results. Therefore, if you draw a, the first result should be a, then o, then e, something like this.

I know little about neural networks. What kind of data will I need to transfer to NN. Maybe this is the x / y coordinate array where the user drew on the pad. Or what type of data is a neural network that expects or provides the best results for handwriting?

Answer 1

Typically, simple NNs for pattern recognition / handwriting accept a 2-bit logic matrix as input; that is, a black and white raster image. Make sure you have a set of tutorials; or let the user learn the algorithm through online learning backprop.

The @FrustratedWithFormsDesigner suggestion of sending an order might make NN smarter, but if you're just learning, first try the version of the bitmap and see how well it works. Also, play with the granularity of the bitmap. Perhaps first try to recognize the number, there are standard data sets for this problem on the Internet.

Answer 2

You not only need to send the X / Y coordinates, but also the ORDER into which they were drawn. Thus, a path can be better than just a set of points. A neural network should be able to handle this, and there are many ways that this might. One way could be to divide the path into n segments for n neurons and each neuron will recognize a piece of letter.

Answer 3

The main process consists in accumulating a number of examples of each identifiable letter, preprocessing the raw data, preparing a collection of candidate models and selecting the final model based on the test results on a separate, stored data set.

The nature of the preprocessing will depend on the data you collect. If this data is "move points", then it is easiest to divide the image into regions and summarize the number of points per region. If you are recording a bitmap image instead, it will be useful to use other pre-processing, such as simple statistics, as well as vertical and horizontal profiles (middle rows and columns).

"Dr. Dobb Journd held a fingerprint recognition contest (using electronic ink data) several years ago. You can read about it here:

http://www.drdobbs.com/184408743;jsessionid=IG5ALGCW1HZZVQE1GHPCKH4ATMY32JVN?pgno=4

... and here:

http://www.drdobbs.com/184408923;jsessionid=IG5ALGCW1HZZVQE1GHPCKH4ATMY32JVN?pgno=2