Abstract: As a nonparametric statistical model, Dirichlet process mixture (DPM) model can be effectively applied to unsupervisedclassification of SAR images. In this paper, an automatic target segmentation on SAR image of MSTAR SAR chips is proposed.First, the number of segments in the image is automatically inferred by using DPM. Then, Markov random field (MRF) is used fordescribing the spatial contextual information of different segments in the image. The final segmentation result is obtained by combining an optimized algorithm of label cost. According to our approach, the number of segments does not need to be predefined.Meanwhile, the reasonable and consistent segmentation results are achieved. The effectiveness of the proposed method is demonstrated by segmentation experiments on MSTAR SAR data sets.