Automatic Determination of Number of Clusters in 1-D Self-Organizing Maps
Cao Lin-ping · Fire Control and Command Control · 2004
One of the problems of Self-Organizing Maps in color segmentation is that the numbers of clusters should be specified in advance and the success of the clustering algorithm largely depends on the specified number of clusters.It's very difficult to determinate the value in learning process without external setup a prior.An approach that can calculate automatically the numbers is suggested in color segmentation.Using weight vectors to represent corresponding classes will lead to some error.Obviously the sum of squared error will decrease monotonically while numbers of output neurons increasing,but its decreasing rate is different.When the decreasing rate begins to get slower substantially,the numbers of classes should be appropriate.The results of experiment validate the validity of the approach.