1998.

abstract bibtex

abstract bibtex

We consider a non-symmetric half plane autoregressive image, where the image intensity of a point is a linear combination of the intensitites of the eight nearest points located on one quadrant of the coordinate plane, plus a normal white noise innovations process. Two types of contaminations are considered. Innovation outliers, where a fraction of innovations are corrupted with a heavy tailed outlier generatioin process, and additive outliers, where a fraction of observations are corrupted. We develop a GM-estimator for the robust estimation of parameters of a contamined autoregressive image model, based on time series GM-estimators introduced by Denby and Martin (1979) applied to the restoration of radar generated images. Ordinary least-squares estimators are asymptotically efficient with a non- contamined gaussian process, like the one considered here. M-estimators behave better when innovation outliers are present, but are very sensitive to additive outliers. A simulation study is carried out, which shows that the GM-estimator introduced here has a better performance with an additive outlier contamined image model than M-estimators and ordinary least squares estimators.

@book{0032461523, abstract = "We consider a non-symmetric half plane autoregressive image, where the image intensity of a point is a linear combination of the intensitites of the eight nearest points located on one quadrant of the coordinate plane, plus a normal white noise innovations process. Two types of contaminations are considered. Innovation outliers, where a fraction of innovations are corrupted with a heavy tailed outlier generatioin process, and additive outliers, where a fraction of observations are corrupted. We develop a GM-estimator for the robust estimation of parameters of a contamined autoregressive image model, based on time series GM-estimators introduced by Denby and Martin (1979) applied to the restoration of radar generated images. Ordinary least-squares estimators are asymptotically efficient with a non- contamined gaussian process, like the one considered here. M-estimators behave better when innovation outliers are present, but are very sensitive to additive outliers. A simulation study is carried out, which shows that the GM-estimator introduced here has a better performance with an additive outlier contamined image model than M-estimators and ordinary least squares estimators.", year = "1998", title = "Digital image restoration using autoregressive time series type models", pages = "53-59", booktitle = "Second Latino-American seminar on radar remote sensing image processing techniques" }

Downloads: 0