Multiprimitive Texture Analysis Using Cluster Analysis and Size Density Function

A novel method of the multiprimitive texture analysis is proposed. This method segments a texture using the watershed algorithm into fragments each of which contains one grain. The size density function of each fragment is calculated, and the fragments are located in the feature space each of whose basis is the size density of a size. The shape of each grain is distorted by the segmentation if the grains overlap, and the watershed algorithm may cause the over-segmentation. Thus the fragments containing grains corresponding to one primitive scatter in the feature space. However, the following cluster analysis collects neighborhood fragments in the feature space into a cluster. The grains in a cluster are regarded as corresponding to one primitive. The number of distinctive primitives shapes is obtained as the number of distinctive clusters, and each primitive is obtained as the central fragment of each cluster.