Multiresolutional regularization of local linear regression over adaptive neighborhoods for color management

A multiresolutional regularization method is proposed for local linear regression that regularizes local mean squared-error by the mean squared-error of a larger neighborhood. The approach is similar in motivation to generalized Tikhonov regularization, but because the regularization trades-off between two like quantities, it is easier to interpret and specify the regularization parameter. Color management experiments with printers are used to compare the proposed regularized local linear regression to ridge regularization and generalized Tikhonov regularization. The local linear regressions use previously-validated adaptive neighborhoods. Results show that significant reductions in error can be achieved over the state-of-the-art.