基于多尺度核模糊聚类的图像分割算法

Image segmentation algorithm based on multi-scale kernel fuzzy cluster

针对现有的模糊聚类算法容易受到噪声影响,难以实现精准分割的问题,提出一种基于多尺度核模糊聚类的图像分割算法。首先,利用高斯滤波对分割图像进行多次模糊,构建多尺度空间。然后,利用图像的局部空间信息和权重信息来修改局部模糊因子,同时引入核函数,用内核诱导距离替换原始欧式度量,增加了其对噪声点和离群值的鲁棒性。将全局隶属度和局部隶属度加权,二次修正隶属度的划分。最后,使用滤波后的上层图像聚类结果依次指导下层图像聚类,避免了随机初始化,有效抑制噪声,提升算法的分割性能。为验证算法的有效性,与其他8种聚类算法进行对比分析,结果表明:所提算法在噪声污染和复杂图像中能够取得较好的分割结果。

As the existing fuzzy clustering algorithms are easily affected by noise,it is difficult to achieve precise segmentation.This paper proposes an image segmentation based on multi-scale kernel fuzzy clustering algorithm.First,Gaussian filter is employed to blur the segmented images many times to build a multi-scale space.Next,the local spatial and weight information of the image is utilized to modify the local blur factors.In the meantime,a kernel function is introduced to replace the original Euclidean metric with the kernel-induced distance,which increases its robustness to noise points and outliers.Then,the global membership and local membership are weighted to modify the division of membership twice.Finally,the filtered upper image clustering results are used to guide the lower image clustering in turn,which avoids random initialization and effectively suppresses noise to improve the segmentation performance of the algorithm.To verify the effectiveness of the proposed algorithm,a comparative analysis with other eight clustering algorithms show that the proposed one achieves better segmentation results with noisy and complex images.