CRIM-FCHO: SAR Image Two-Stage Segmentation With Multifeature Ensemble

This paper investigates the synthetic aperture radar (SAR) image segmentation in terms of feature analysis and fusion and develops a new algorithm based on multifeature ensemble accordingly. This paper is characterized by two aspects. First, multiple heterogeneous features are extracted to accurately describe the objects in SAR images. These features are then integrated in the feature level and the similarity level, respectively, to avoid the mutual influences between different kinds of features and maximize the discriminability of the similarity measure between objects. Second, a two-stage algorithm consisting of a coarse merging stage and a fine classification stage is proposed. In the coarse merging stage, a context-based region iterative merging algorithm is designed to merge most of the unambiguous superpixels in image domain at a high speed. In the fine classification stage, a fuzzy clustering algorithm incorporating hybrid optimization is developed to balance the efficiency and the robustness of the algorithm by simultaneously searching heuristically in the complete high-dimension feature space and searching along the direction of the gradient steepest descent in each feature subspace. The effectiveness of the proposed method has been successfully validated on synthetic and real SAR images.