多特征-谱聚类的SAR图像溢油分割

SAR oil spill image segmentation based on a multi-spectral clustering algorithm

经典的K聚类算法,并不适合实现任意形状的聚类,而且有容易陷入局部最小值的不足.提出基于多个纹理特征的谱聚类算法,该方法用灰度共生矩阵(GLCM)提取合成孔径雷达(SAR)图像的多个特征值,构建谱聚类的特征矩阵,并依据规范切准则,用K均值聚类的方法对拉普拉斯矩阵的第二小的特征值对应的特征向量进行聚类,实现基于SAR图像的溢油的分割.新方法与传统的K聚类方法比较,可以减少相干斑噪声对分割结果的影响,较好的保持图像边缘.仿真结果显示,该算法对于相干斑噪声影响较大的图像具有较强的鲁棒性.

The classic K-mean clustering algorithm is not suitable for the circumstances of arbitrary shapes and is prone to use the local minimum.In order to fix these shortcomings,a spectral clustering algorithm based on multi-texture characteristics was proposed.The algorithm first used gray-level co-occurrence matrices（GLCM） to extract three features of the synthetic aperture radar（SAR） image and construct a characteristic matrix of spectral clustering.Next using the N-cut（Normalized-cut） criterion,it clustered the eigenvector corresponding to the second small eigenvalue of the Laplacian matrix in order to carry out the SAR oil spill image segmentation.Compared with the classic K-mean algorithm,the proposed method reduces the influence of coherent scattering noise on the segmentation result and efficiently conserves the edge of the image.The simulation results also show that the new method has a strong robustness for an image badly affected by the coherent scattering.