Abstract: This paper presents a rotation-invariant approach for segmenting textured images. The textured images are modeled as three hierarchical processes, namely, the region process, the texture process, and the degradation process. The whole procedure includes extraction of rotation-invariant features, smoothing of these features, initial segmentation based on three-layered neural networks, and extensive segmentation using multi-valued Boltzmann machines. Experiments on natural textures have shown its efficiency.