基于分水岭-谱聚类的SAR图像分割

SAR IMAGE SEGMENTATION BASED ON WATERSHED AND SPECTRAL CLUSTERING

由于谱聚类是基于图论的、以相似性为基础的聚类方法,需要计算图像中每对像素点之间的相似性.当图像很大时,计算相似性矩阵和求解相应的特征值、特征向量是很困难和耗时的.为此,针对合成孔径雷达(SAR)图像的特点,提出了一个两阶段的图像分割方法,首先采用分水岭算法对图像进行过分割,然后再用改进的谱聚类算法进行聚类.新方法不仅可以减少噪声对分割结果的影响,很好地保持图像边缘,而且对时间要求较高的应用也具有一定的参考价值.为了验证新方法的有效性,将其用于SAR图像分割,取得了较优的分割结果.

Spectral clustering is a similarity-based clustering algorithm on graph theory. When the processed image is very huge, it is very difficult and time-consuming to compute affinity matrix and its eigenvalues and eigenvectors. Aiming at the characteristics of synthetic aperture radar（SAR） images, a two-stage image segmentation algorithm was proposed, in which watershed was used to produce over-segmentation and an improved spectral clustering algorithm was applied to perform final clustering. The new algorithm can not only reduce the noise in SAR images and keep their boundary very well but also is valuable to the application of higher demand for time. To verify the performance of the proposed algorithm, it was applied to segment SAR images, and better segmentation results were obtained.