Natural texture images can exhibit high intra-class diversity due to the acquisition conditions. To reduce its impact on classifi cation performances, the geometry of the cluster in the feature space should be considered. We introduce the Spherically Invariant Random Vector (SIRV) representation, which is based on scale-space decomposition, for the modeling of spatial dependencies characterizing the texture image. From the speci c properties of the SIRV process, i.e. the independence between the two sub-processes of the compound model, we derive a centroid estimation scheme from a pseudo-distance i.e. the Jeff rey divergence. Next, a K-centroids based (K-CB) supervised classifi cation algorithm is introduced to handle the intra-class variability of texture images in the feature space. A comparative study on various conventional texture databases is conducted and reveals the impact of the proposed classi cation algorithm.