Unsupervised image segmentation based on Nonsubsampled Contourlet hidden Markov trees model

Based on the shift invariance and multidirectional expansion properties of nonsubsampled contourlet transform, a new image segmentation combining hidden Markov trees model with Bayesian approaches is proposed here. The training blocks can be got in pre-segmentation by using histogram approximation. To integrated use the information of different scales, we use the contextual model to segment raw segmentations and fuse them to get the final image. We compared our result with wavelet domain HMTseg method and contourlet domain HMTseg method, the result shows that our method has better performance in edges but lower missed classed probability. The simulation results of synthetic mosaic image, aerial image and SAR image are showed to prove the generalization of this method.