We propose a new class of non-linear image recovery operators based on local smoothing applied both in the spatial domain (horizontal smoothing) and in the gray-level domain (vertical smoothing). The commonly used horizontal smoothing gives the way of reducing the noise level (noise filtering), whereas the vertical smoothing yields the edge preservation property. The method is indexed by a real parameter yielding a generalized class of local harmonic filters. In particular the classical Min/Max filters as well as all linear locally weighted filters can be obtained as special cases.