基于轮廓波变换和改进模糊c均值聚类的红外图像分割  

Contourlet transform and improved fuzzy c-means clustering based infrared image segmentation

作　　者：刘刚[1] 梁晓庚[1,2] 张京国[1,2] 

机构地区：[1]西北工业大学自动化学院,陕西西安710072 [2]洛阳光电技术发展中心,河南洛阳471009

出　　处：《系统工程与电子技术》2011年第2期443-448,共6页Systems Engineering and Electronics

摘　　要：针对红外图像分辨率低、对比度弱、易受噪声污染等特点,给出了轮廓波变换与模糊c均值聚类相结合的红外图像分割方法。该方法首先在假定图像轮廓波变换系数的先验为高斯分布的基础上,将基于最大后验概率准则的比例萎缩法应用于红外图像降噪,以完成分割前的预处理过程,然后利用改进的模糊c均值算法对降噪后的红外图像进行分割。针对标准模糊c均值分割存在的问题,本文提出的方法从基于样本直方图的最小最大距离法的初始聚类中心确定、考虑邻域像素相关性的样本点聚类权值和邻域隶属度修正三个方面加以改进,在保证分割精度的基础上,进一步滤除降噪阶段遗留下的图像噪声。对一系列红外图像进行实验的结果表明,相对于标准模糊c均值算法,本文提出的改进算法划分熵平均降低约10%,区域对比度提高约27%,能够实现对受到噪声污染红外图像的有效分割。Aiming at the feature of low resolution and faint contrast for infrared images,a segmentation algorithm is presented based on the contourlet transform and fuzzy c-means clustering.This method assumes that the prior distribution of the original image coefficients and the noise coefficients both are Gaussian and applies the proportional shrinkage based on the rule of maximum a posteriori to the infrared image＇s denoising in contourlet domain.Subsequently,an improved fuzzy c-means clustering algorithm is put forward to segment the denoised image.This method improves the segmented performance in three aspects which are the method of the minimum-maximum distance based on the histogram to compute the original clustering center,the computing of the sample weight and the revising of the membership grade during the clustering procedure by considering the pixel＇s neighbor.The experimental results show that the proposed method,compared with the standard algorithm,can make the partition entropy decreased by 10% and the region contrast ratio increased by 27%.So this algorithm can segment the infrared image which is polluted by noise effectively.

关 键 词：红外图像分割 轮廓波降噪 最大后验概率 模糊C均值聚类 聚类中心 样本权值 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]