A distribution-free criterion for robust identification, with applications in system modelling and image processing

Abstract A new distribution-free identification criterion based on sign changes in the residuals is presented. The maximization of this non-differentiable criterion is carried out by a global optimization routine using an adaptive random search strategy which is demonstrated to furnish very satisfactory results. The robustness of the resulting identification procedure is tested by treating three examples. The first one is a simulated example, representative of a situation where many outliers are present. The second one is concerned with the estimation of the parameters of a compartmental model routinely used for the functional study of liver and biliary ducts of patients with Tc-99m-diethyl IDA. The last one deals with the problem of automated registration in digitized medical image processing. The new criterion has already proven to yield original and efficient methods for change detection in dissimilar images. These examples form the basis for a comparison of the new estimator with more classical ones.