Even if the Autoregressive (AR) model is widely used in several frameworks such as speech, seismic and biomedical or textured image analysis, it is not really efficient to represent periodical signals. In this paper we propose first to investigate the limits of the autoregressive process and secondly to present an alternative approach based on the Wold decomposition. We then present a comparative study between AR model and Wold decomposition in the framework of the reconstruction of textured images