基于图的加权核K均值的图像多尺度分割

Multiscale Image Segmentation Based on Graph Weighted Kernel K-means

提出改进的最小割(IMC)模型以避免分割出小的孤立点集,研究了改进的最小割模型与加权核K均值之间的等价关系,列举了几种常见的用于建立图割模型边权值的相似度函数,并分析了其对分割结果的影响。在此基础上,设计了一个基于图的加权核K均值图像多尺度分割方法,该方法既避免了基于图割的图像分割中图谱的求解问题,又避免了加权核K均值方法中核矩阵的选取问题,同时实现了对图像多尺度的分割。通过对该方法进行抗噪性能的分析,以及在光学图像上对实验结果进行比较,验证了所提出方法的有效性。

An improved minimum cut model is presented considering that the minimum cut criteria favors cutting small sets of isolated nodes,then equivalence relation between the improved minimum cut model and weighted kernel k-means is researched,and the influence of different similarity functions on the results of segmentation are also analysed. And based on these,a multiscale image segmentation method based on graph weighted kernel k-means is proposed,this method avoids calculating graph spectral,which is a key step when using graph cut model to segment images,also,it avoids selecting kernel matrix,which is important to the weighted kernel k-means,finally it realizes multiscale image segmentation. The analysis of anti-noise and experimental results on a number of optical images show the effectiveness of this method.