A set of multiresolution texture features suitable for unsupervised image segmentation

We propose a set of multiresolution features for texture description. Image smoothing at multiple scales using the fast smoothing B-spline transform is performed and a number of features, such as the local area, the normal vector dispersion and the gradient orientation, are computed from each scale. A simple disparity function is applied to assess the discriminative power of these features with comparison to other texture methods. Being effective even for small observation windows, the proposed features are suitable for high-resolution texture segmentation.