Multidimensional autoregressive parameter estimation using iteratively reweighted least squares

Abstract

Two-dimensional robust autoregressive parameter estimation is performed on image data using an iteratively reweighted least squares (IRLS) procedure which explicitly identifies the model outliers. In practice, these outliers often arise from nonhomogeneous image structures. An initial least median of squares estimate is used to obtain a more robust version of IRLS. Both versions of the IRLS algorithm are tested experimentally on synthetic and real image data. A whiteness measure, based on a two-dimensional version of the Box and Pierce portmanteau test, serves as a useful performance evaluator. The experimental results demonstrate that the robust parameter estimators can offer significant improvement over the classical least-squares estimator on image data that deviates from the autoregressive model. These results have potential applications in image processing, including image coding and object detection. >