基于Contourlet域图谱聚类和多尺度Markov模型的多光谱遥感图像分割

Segmentation of multispectral remote sensing image based on spectral clustering and multiscale Markov model in Contourlet domain

针对多光谱遥感图像的特点,结合图谱聚类、Contourlet系数分布的统计特性和多尺度Markov模型,提出了一种基于Contourlet域图谱聚类和多尺度Markov模型的分割(CSCMMS)方法。首先对待分割图像进行Contourlet变换,利用图谱聚类对最粗尺度低频图像聚类得到可靠的初始分割结果;然后利用互信息构造Contourlet域的多尺度Markov模型,结合多尺度、多方向的图像信息将低频图像的初始分割结果逐尺度传递到最细尺度,得到原始图像的分割。对合成图像和多光谱遥感图像的实验结果表明,提出方法在边缘信息保持和噪声敏感性上具有明显改进,错分率和运算时间进一步降低。

Spectral clustering methods have recently shown great promise for the problem of image seg- mentation. However, the computational demands of these approaches make them infeasible to large problems such as multispectral remote sensing images. According to the particularities of multispectral remote sensing images, a segmentation method is proposed based on spectral clustering, statistics characteristics of Contourlet coefficients and multiscale Markov model. First, the spectral clustering is implemented to the coarsest lowpass image of multispectral remote sensing image in Contourlet domain to obtain the initial segmentation result; Second, a multiscale Markov model, which contains the initial segmentation result as the coarsest scale, is constructed using mutual information to capture the relationship between Contourlet coefficients in the same scale and across scales Third, the final segmentation result is obtained by confusing multiscale and multi-directional image information based on the multiscale Markov model. Compared with the classical spectral clustering method and the multiscale segmentation method based on HMT model in Contourlet domain, the segmentation results for both synthetic images and real multispectral remote sensing images show that the proposed method not only has better performance in edges preservation and noise sensitivity, but also has lower misclassification probability and running time, and it can achieve satisfactory segmentation results.