Unsupervised segmentation of noisy image in a multi-scale framework

We present a multi-scale framework for segmentation of image modeled by a Markov random field (MRF). In this framework, a multi-scale representations of the original image are derived in nonlinear scale-space using anisotropic diffusion, which has the advantage of smoothing unwanted structures while preserving semantically meaningful structures at any scale. Then we apply segmentation using a “from coarse to fine” scheme. A histogram analysis method is developed to approximately estimate the parameters and the maximum a posterior (MAP) estimation of the label field is obtained at the coarsest scale using fast iterative conditional modes (ICM), and then the labeling result is mapped to the next-finer scale taken as the initial labeling, while the parameters is modified using maximum likelihood (ML) estimation. This procedure is continued until the finest scale is reached. At each scale, simple and fast ICM algorithm is applied. Experiment results on real and synthetic image show good performance of our scheme