Edge-Aware Superpixel Generation for SAR Imagery With One Iteration Merging

Most of the existing superpixel generation methods are based on local iterative clustering. However, such methods have the following shortcomings: 1) these methods require several iterations and the number of iterations is difficult to determine and 2) the generated superpixel lacks explicit connectivity without a postprocessing step. Aiming to overcome the limitations, we propose an edge-aware superpixel generation with one iteration merging (ESOM) for synthetic aperture radar (SAR) imagery. In specific, we introduce a ratio-based edge detector with a Gaussian-shaped window to extract the edge information and an edge-aware dissimilarity is defined. Then, a new merging method termed as one iteration merging is proposed, which leverages the continuity of the adjacent pixels and ensures the connectivity of superpixel. Furthermore, instead of iterative clustering, the one iteration merging is achieved in only one iteration without determining the number of iterations and hence efficient in computation. Experiments on two real SAR images demonstrate that the proposed method yields substantially better performance than some state-of-the-art methods.