An approach for determining optimal linear image restoration filters is developed by the authors in great detail. A subspace information criterion (SIC), which is an unbiased estimator of the expected squared error with finite samples, is employed. When using an SIC, it is assumed that a linear filter is available that provides an unbiased estimate for the image, and that the noise covariance operator is known.

A procedure using a generalized inverse and a wavelet transform for estimating these values in practice is described. An analytic optimal parameter derivation using SIC for a moving average filter is included, along with experimental results. The mathematical concepts are clearly explained, and further clarified with figures.

A major part of the work presented consists of a set of experiments on images degraded with a horizontal blur and normally distributed noise. The SIC optimal regularization filter (optimized over a finite set of regularization parameters) is compared with a set of six filter optimizations from the literature, under different degrees of image degradation. The results are well organized into a combination of graphs and images that demonstrate the utility of SIC. An extensive list of references is provided, and several suggestions are made for further work.