Unsupervised SAR image segmentation using gradient Triplet Markov fields model

Abstract

For image segmentation, methods based on Markov random field (MRF) have been widely used due to the capability of incorporating contextual information into account. Among these methods, the recently proposed Triplet Markov fields (TMF) model enhances the image nonstationary modeling by introducing an auxiliary field. According to the introduced auxiliary field, TMF can partition the nonstationary image into several different stationary parts, and then models each stationary part using specific pattern of local interactions. However, synthetic aperture radar (SAR) images generally do not contain obvious repetitive different local image patterns, and it makes the TMF model difficult to explicitly decide the number of different stationary parts. In this paper, we propose a segmentation method for SAR images using gradient TMF model (GTMF). Compared to the TMF model, the auxiliary field in GTMF is redefined to indicate the dominant direction of local image contents, which gives explicit nonstationary divisions of SAR images. Based on the local directional nonstationary division, the posterior distribution is constructed with local contextual information adaptively emphasized along the dominant direction within each local neighborhood. The Bayesian MPM inference method can be applied to give the segmentation result. Multiple real SAR images are utilized in the experimental analysis, and the effectiveness of the proposed method is validated accordingly.