A new constrained spectral clustering for SAR image segmentation

Pairwise constraints have been successful to be applied in traditional clustering methods. However, little progress has been made in incorporating them into spectral clustering. In this paper, we propose a new method to combine pairwise constraints with spectral clustering and apply it to SAR image segmentation. Firstly, we learn a distance metric using pairwise constraints. In doing so, an affinity matrix is obtained by the Gaussian function on the learned distance metric. Then we perform the spectral decomposition on the affinity matrix and we get the spectral features of data points. Finally, the constrained k-means is used to cluster spectral features instead of k-means used commonly. We apply the proposed method to synthetic aperture radar (SAR) image segmentation. Experimental results show that it is effective for SAR image segmentation.