Dhivehi Character Recognition
Javascript+DHTML pattern recognition implementation 2.1
by Jaa ( http://jaa.technova.com.mv/ )
, 2005


Introduction:

This is just a fun little program to show an implementation of rudimentary character recognition for Thaana characters.

Technical Notes:
The entire implementation is done using JavaScript and DHTML. You are welcome to study the code to see how it works. The code is well commented and maybe a good starter into AI and pattern recognition basics. It uses a single layer single perceptron model to really simplify things however is a good enough practical implementation to work for characters drawn on a 10x10 grid. The grid makes up the input data to the neural network. The neural network is hardcoded into the page and has definitions for each character in the alphabet.

It is error prone due to rigid grid input for characters it uses. It also imposes limitations where by characters are required to be drawn to fit inside the box perfectly. The current version (2.0) does not use biases for the summing function thus tends to throw off recognition processes and a lot times ends up defaulting to "haa"
as a result. The neural nets also need more training as currently, it has only been under two passes of training per character.

How to use it:
1. Click and hold the left mouse button and draw a Thaana "akuru" in the box below. You are free to draw the character as you please but try drawing big in order to fit the character to the size of the box.
2. Then click "Recognise" button.
3. Hit clear to draw another character.

Requirements:
+ Firefox or Internet Explorer 4+ or Opera

Changelog:
v2.0 ( 20-11-05 ):
- Reworked code from first version released in 2003.
- New interface and code now commented.
- Cross browser compatibility achieved.

v2.1 ( 21-11-05 ):
- Fixed major flaw in the reworked code v2.0

 


Draw Thaana alphabet: