Partitional Clustering Techniques for Multi-Spectral Image Segmentation

Analyzing unknown data sets such as multispectral images often requires unsupervised techniques. Data clustering is a well known and widely used approach in such cases. Multi-spectral image segmentation requires pixel classification according to a similarity criterion. For this particular data, partitional clustering seems to be more appropriate. Classical K-means algorithm has important drawbacks with regard to the number and the shape of clusters. Probability density function based methods overcome these drawbacks and are investigated in this paper. Two main steps in data clustering are dimension reduction and data representation. Methods like PCA and ICA often perform dimension reduction step. To achieve a complete and more reliable representation of the data, a magnitude-shape representation is described, it takes into account both the magnitude and shape similarities between pixels vectors. The bases of PCA and magnitude-shape representation are explored to highlight the main differences and the advantages of our method over PCA. Experimental results confirm that this method is a reliable alternative to classical linear projection methods for dimension reduction. a)