This paper adresses the problem of modeling object segmentation in the visual cortex using oscillations. The proposed architecture is based on a network of locally connected FitzHugh-Nagumo oscillators which receive graded external input. We show the suitability of such a network to encode the stimulus since the amplitude of oscillations increases monotonically as a function of the input in the neighborhood of a bifurcation, while the frequency remains nearly constant. However, due to the diffusive effects of the Laplacian connectivity, the oscillators tend to be in phase even when they represent different objects. Therefore a desynchronization mechanism, which represents spatial information about the objects, is added. The overall dynamics are described and simulation results on real images are shown.