This paper deals with the problem of recovering a sparse unknown signal from a set of observations. The latter are obtained by convolution of the original signal and corruption with additive noise. We tackle the problem by minimizing a least-squares fit criterion penalized by a Geman-McClure like potential. The resulting criterion is a rational function, which makes it possible to formulate its minimization as a generalized problem of moments for which a hierarchy of semidefinite programming relaxations can be proposed. These convex relaxations yield a monotone sequence of values which converges to the global optimum. To overcome the computational limitations due to the large number of involved variables, a stochastic block-coordinate descent method is proposed. The algorithm has been implemented and shows promising results