Recognition System
of Japanese Characters
Self-Organization Map
The Self-Organization Map (SOM) is a competitive network used to classify input vectors. The principal elements of a SOM are:
Input vector (p).
Weight Matrix (IW).
Competitive Function (S).
The SOM obtain a negative distance between the input vector and the weight matrix. The competitive function gives a output one to the neuron whose distance is the less negative value. All others are zero.
To transform from continous data to discrete data, we trained the SOM as shown below. A matrix of points was defined using a gridtop topology. This network is a bidimensiona array (4x4) of neurons.
The HMM_Evaluate module calls a S-Function to evaluate the symbol sequence data. To pursuit a partial recognition a buffer (25 data samples) is used.
The first figure below, shows the input that is used by the SOM. The output is shown in the second figure. This data is used by the HMM_Evaluate module.
To test this system, we presented 20 samples of each model trained (off-line). Two tests were done: full trajectory (left table) and partial trajectory (right table).
Last update 01.15.2004
These pages are maintained by Jorge Solis