SOS Boosting for Image Deblurring Algorithms

Shahar Romem Peled, Yaniv Romano, Michael Elad

Abstract: The work reported by Romano and Elad presents the SOS boosting - a generic recursive method for improving image denoising algorithms. Appealing properties of the SOS scheme are its flexibility to work with any denoising method, its simplicity, and its robustness. In this article we aim to generalize this to the image deblurring task. The proposed SOS procedure for deblurring is given by the following iterative steps: [S]trengthen the signal by blurring the output of the previous iteration and adding it to the blurred input image, [O]perate the deblurring algorithm on the strengthened image, and [S]ubtract the previous output from the current one. We demonstrate this procedure for several state-of-the-art methods (BM3D, EPLL and NCSR), showing the potential gain in output quality for each. As in the original SOS, we manipulate the iterative algorithm by two parameters, better controlling its steady state and rate of convergence.