High-order CRF based on product-of-experts for unsupervised SAR image multiclass segmentation

Conditional random fields (CRF) model is suitable for image segmentation because this model directly defines the posterior distribution as a Gibbs field and allows one to capture the dependencies of the observed data. However, this model has a limited ability to capture the high-order label dependencies with only the pairwise potential being constructed generally. Moreover, for synthetic aperture radar (SAR) image segmentation, SAR scattering statistics that are essential to SAR image processing are not considered in CRF model. Then in this paper, we propose a high-order CRF model based on product-of-experts (POE) for unsupervised SAR image multiclass segmentation, named as HCRF-POE model. HCRF-POE model decomposes the high-order label dependencies into the low-order ones and constructs the high-order potential based on POE, thus effectively capturing high-order label dependencies. In addition, to capture SAR data information including the textural features and the scattering statistics, in a more completed manner, HCRF-POE model integrates SAR data information under unsupervised Bayesian framework. The effectiveness of HCRF-POE model is demonstrated by the application to the unsupervised segmentation of the simulated image and the real SAR images.