We describe a method for performing deghosting in the image domain. The method consists of a multidimensional deterministic deconvolution of a standard migrated image using non-Hessian form operators. The operators, created by modelling with ghost effects followed by standard migration, compensate for both dip-dependent illumination and ghost effects, and provide an estimation of the reflectivity. Applying a standard modelling then migration operator to the estimated reflectivity yields the 3D deghosted image. The operators are approximated with a set of filters evaluated in the image domain also known as Point Spread Functions (PSFs). We also argue that the method is well suited to complement any data domain (pre-stack) deghosting performed early in the processing chain in order to facilitate subsequent processing. The method is illustrated on complex synthetic data with significant geometry irregularities and on a data example from the North Sea.