Semantic labelling of SAR images with conditional random fields on region adjacency graph

Abstract

Scene segmentation and semantic labelling are important for analysing and understanding synthetic aperture radar (SAR) images. In this study, the authors propose an effective and efficient labelling method for SAR images with conditional random fields on a region adjacency graph (CRF-RAG). More precisely, for an SAR image, a region adjacency graph (RAG) representation is firstly built on an initially over-segmentation of the image. Subsequently, a conditional random field (CRF) model is established over the RAG instead of over pixels. To train and infer the CRF-RAG model, a fast max-margin training strategy and the graph cut optimisation method are finally employed. As the CRF model is based on RAG, the computation complexity of the model can be reduced significantly. Compared to the Markov random field (MRF) model on RAG, the proposed CRF-RAG model is more efficient to incorporate different measures of SAR images, such as scattering intensity, texture and image context, into a unified model. Experiments on the TerraSAR-X imagery achieve promising results with modest computation cost, which validates the generality and flexibility of the proposed method.