In this paper an unsupervised image segmentation method is presented, which combines wavelet domain Markov random field (WD-MRF) with the modified fuzzy c-means (FCM) clustering algorithm. At the label establishment stage, a WD-MRF tree is employed to model the statistical properties of multiresolution wavelet coefficients. Each wavelet coefficient is characterized by a feature field and a label field model, the feature field being regarded as an observation of its label field, and the label indicating that the wavelet coefficient belongs to a region. After the model parameters are estimated by expectation maximization algorithm, all regions of image are labeled by maximum a posterior principle. At the original image segmentation stage, the contents of the image are formulated as a fuzzy objective function, where the persistence of interscale wavelet coefficients is considered, and by minimizing the objective function, the novel fuzzy segmentation algorithm is derived. The experiments with synthetic images are carried out, and the results show that the proposed method outperforms conventional FCM and fixed resolution Bayesian segmentation algorithm, such as accurately locating image edges, correctly identifying different regions.