This paper integratesMarkov random fields (MRFs) with type-2 fuzzy sets (T2 FSs) referred to as T2 FMRFs, which can handle the fuzziness of the labeling space as well as the randomness of observations within the unified framework. Because fuzzy and random uncertainties exist in many computer vision problems, we extend the maximum a posteriori (MAP) criterion for the best labeling configuration by T2 FSs operations. We apply T2 FMRFs as character models to similar handwritten Chinese character recognition on ETL- 9B and KAIST databases. Experimental results show that T2 FMRFs have a better classification and generalization ability for similar patterns than classical MRFs.