In this paper, joint conditional probability is localized to better capture the local properties of a neighborhood for image segmentation. A new local spatial likelihood is defined for a neighborhood, which gives rise to local spatial posterior associated with the defined local prior. The proposed method associates a novel nonparametric approach for estimating the underlying distributions and is compared with a parametric approach. Both approaches segment images by maximizing the local spatial posterior function. The results indicate that the spatially localized posterior function overcomes the inherent errors of general posterior function and gives remarkable robustness against heavy noises.