Automatically enumerating image data clusters using pixel co-density
Ryan A. Mercovich · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2012
Typical automatic clustering methods struggle to determine the correct number of clusters to properly characterize the data. To estimate the number of clusters in a spectral image data cloud explicitly from the data structure, the pairwise relationships between pixels in the n-dimensional spectral space are exploited. By plotting the average ith co-density between pixels and neighbors, a monotonically increasing function will emerge that characterizes the clusters in the data. Large upward steps in the average neighbor distance function represent the well-grouped clusters in the data. This process can accurately identify the number of clusters in a wide variety of image data automatically.