The problem of recognizing rotated homogeneous textured images is addressed. The aim is to provide some comparison results between two classical non-parametric techniques — namely Zernike moments and Fourier-Mellin descriptors — and a new parametric approach involving the Wold decomposition of 1-D processes. In order to obtain translation invariance, all these methods start with the computation of the 2-D normalized autocovariance of textures. The techniques and numerical aspects of the computation of invariant features are briefly described. Experiments performed on a texture database show that the parametric model provides encouraging recognition rates comparable with the Zernike moments, along with the important advantage of its parcimony w.r.t. the classical approaches.