Comparison of Two Unsupervised Methods of Classification for Segmenting Multi-Spectral Images

Some clustering algorithms require assumptions (such as number and shape of classes), which limit their performances or provide wrong results. On the contrary, methods based on the estimation of the probability density function (pdf) do not make any assumption neither on the classes shape nor on their number. Two methods based on the pdf, are explored and applied to the segmentation of a multi-spectral image of a cereal grain. The first one is inspired from the estimation of the pdf Parzen-Rosenblatt and the second one estimates the support of the pdf through the support vector theory