I. Introduction
Texture characterization and segmentation are long-standing problems in Image Processing that received significant research efforts in the past and still attract considerable attention. In essence, texture consists of a perceptual attribute. It has thus no unique formal definition and has been envisaged using several different mathematical models, mostly relying on the definition and classification of either geometrical or statistical features, or primitives (cf., e.g., [1], [2] and references therein for reviews). Texture analysis can be performed using either parametric models (ARMA [3], Markov [4], Wold [5]) or non parametric approaches (e.g., time-frequency and Gabor distributions [6], [7]). Amongst this later class, multiscale representations (wavelets, contourlet, …) have repeatedly been reported as central in the last two decades (cf., e.g., [8]–[11]). They notably showed significant relevance for the large class of scale-free textures, often well accounted for by the celebrated fractional Brownian motion (fBm) model, on which the present contribution focuses. Examples of such textures are illustrated in Fig. 1(a). While scale free-like texture segmentation has been mainly conducted by exploiting the statistical distribution of the pixel amplitudes [12]–[14], the present paper investigates the relevance of local regularity-based analysis.