用于图像分割的非局部空间约束的核FCM算法

Kernelized FCM Algorithm with Non-Local Spatial Constraint for Image Segmentation

图像分割是图像分析、图像理解的前提和关键,其结果直接决定着图像分析和理解的质量.模糊C均值(Fuzzy C-Means,FCM)聚类算法是一种常用的图像分割算法.然而,由于经典的FCM算法只考虑像素自身,从而对外围噪声比较敏感.因此,提出了一种改进的用于图像分割的FCM聚类算法.该算法通过利用核方法修改FCA-NLASC算法中的目标函数而实现,即用核距离替代FCA-NLASC中的欧氏距离,相应地得到核FCA-NLASC聚类算法——KNLASC-FCM聚类算法.利用提出的算法分别进行人工合成图像和实际图像的实验结果表明,当图像含有噪声时,与算法FCA-NLASC相比,KNLASC-FCM算法在主观视觉、客观量化两方面的评价中均具有更好的分割性能.

Image segmentation is the crucial and first step of image analysis and understanding,and the segmenting result directly determines the quality of image analysis and understanding.Fuzzy C-means(FCM)is a commonly used algorithm for image segmentation.However,conventional FCM algorithm is sensitive to noise due to not taking neighboring pixels into consideration.To this end,a modified FCM clustering algorithm for image segmentation is proposed in this paper.The algorithm is realized by modifying the objective function in the FCA-NLASC algorithm using kernel method,i.e.,the original Euclidean distance in the FCA-NLASC is replaced by a kernel-induced distance,and thus corresponding algorithm is derived and called as the kernelized FCA-NLASC,shorted in KNLASC-FCM clustering algorithm.Experimental results on both artificially synthesized image and real image demonstrate that when images are contaminated by noises,the KNLASC-FCM algorithm has better segmentation performance than FCA-NLASC both in visually subjective and quantitatively objective evaluations.