Mode boundary detection by relaxation for cluster analysis

摘要

In cluster analysis, modes can be identified using differential operators to detect their boundaries. Raw boundary detector output is often quite noisy so that it may be difficult to obtain good results from edge following techniques.In such situations, a relaxation labeling process can be used to iteratively reinforce the mode boundaries detected by a differencing operation.Results of some simulations presented demonstrate that points lying on mode boundaries reach high boundary probabilities, while other points attain high no-boundary probabilities, so that thresholding becomes trivial.Examples of the performance of a clustering scheme based on the so-obtained mode boundaries are given using artificially generated data sets.