The model of Gibbs random fields is widely applied to Bayesian segmentation due to its best property of describing the spatial constraint information. However, the general segmentation methods, whose model is defined only on hard levels but not on fuzzy set, may come across a lot of difficulties, e.g., getting the unexpected results or even nothing, especially when the blurred or degraded images are considered. In this paper, two multiclass approaches, based on the model of piecewise fuzzy Gibbs random fields (PFGRF) and that of generalized fuzzy Gibbs random fields (GFGRF) respectively, are presented to address these difficulties. In our experiments, both magnetic resonance image and simulated image are implemented with the two approaches mentioned above and the classical "hard" one. These three different results show that the approach of GFGRF is an efficient and unsupervised technique, which can automatically and optimally segment the images to be finer.