Mean Shift Particle-Based Texture Granularity

This paper defines a new measure of texture granularity based on image particles obtained via the mean shift segmentation algorithm. A given texture is segmented a number of times to produce a scale-space hierarchy of increasingly finer images. In each instance, the segmented texture of the previous layer is used as a mask in the segmentation of the current layer. In this way, finer regions are forced to become nested within the larger enclosing regions of the previous layer. Next, the numerical differences between each layer and its predecessor are computed. Empirical evidence indicates that the most accurate estimate of the actual texture particle size occurs between the two layers of greatest numerical difference. Using this difference information, the overall granularity of the texture is calculated. Unlike existing measures of granularity which focus mainly on gray-level variations, this measure describes a texture's granularity in terms of its intrinsic micro structure. It is expected that this interpretation of granularity will more closely approximate that of the human visual system.