Combining regularization frameworks for image deblurring: optimization of combined hyper-parameters

Regularization is an important tool for restoration of images from noisy and blurred data. In this paper, we present a novel regularization technique (CGTik) that augments the conjugate gradient least-square (CGLS) algorithm with Tikhonov-like prior information term. This technique requires the appropriate selection of two hyper-parameters, the number of iterations (N) and amount of regularization (a). A method to select good values for these parameters is developed based on the L-curve technique. Tests were performed by calculating reconstructed images for each algorithm for heavily blurred images. CGTik showed improved restored images compared to the separate Tikhonov and CGLS algorithms.