GCV and ML Methods of Determining Parameters in Image Restoration by Regularization: Fast Computation in the Spatial Domain and Experimental Comparison

Abstract Many linear image restoration methods minimize a compound criterion which balances some fidelity to the observed data via a least-squares measure, and some fidelity to prior information on the unknown object via a smoothing function. In the case of quadratic criteria, this regularization scheme can be interpreted as a Bayesian estimation of an object which is modeled as a Gaussian random field. The global degree of regularization is controlled by a scalar smoothing parameter and by the prior covariance matrix of the object. These quantities are usually unknown and should also be determined from the observed image. Two ways of choosing these parameters are discussed and compared on both synthetic and real images: maximum likelihood (ML) and generalized cross-validation (GCV). Particular attention is paid to implementation problems. Both criteria are evaluated in the spatial domain using fast pixel-recursive techniques. Results show that GCV is more robust than ML with respect to modeling assumptions, and should therefore be preferred in real-world applications.