A practical stopping rule for iterative signal restoration

Iterative signal restoration is a common and simple approach for recovering degraded signals. Unfortunately, iterative algorithms often converge slowly, and the convergence point is usually not the best restoration because of noise amplification. Therefore, a stopping rule should be imposed to maximize the effectiveness of the iterative algorithm. We demonstrate the value of randomized generalized cross-validation (RGCV) as a stopping rule for linear iterative restoration algorithms with simple initial conditions and show that it can be used on relatively small data sets with confidence. We also illustrate the performance of the RGCV criterion for various degrees of blurring and noise