Impulse noise removal by an adaptive trust-region method

This paper suggests a two-phase scheme for the impulse noise removal. In the first phase, an adaptive median filter (AMF) identifies noise candidates created by salt-and-pepper method. In the second phase, a new trust-region phase recovers noise pixels detected by the first phase. The trust-region phase produces a new adaptive radius strategy using the projected gradient for the cases where iterations are successful and a low-memory nonmonotone spectral projected gradient method (SPG) where iterations are unsuccessful. We solve the trust-region subproblem without bound constraints by the dogleg strategy (DG) using the structure of the compact limited memory Broyden-Fletcher-Goldfarb-Shanno technique. The solution is then projected into the boundary region. Numerical results are given to illustrate the efficiency of the new approach for the impulse noise removal.