多尺度无监督彩色图像分割  

Multi-scale Unsupervised Color Image Segmentation

作　　者：陈志刚[1] 陈爱华[1] 崔跃利[1] 项美晶[1] 

机构地区：[1]台州学院物电学院,浙江台州318000

出　　处：《光子学报》2011年第10期1553-1559,共7页Acta Photonica Sinica

基　　金：浙江省自然科学基金(No.Y4110207)资助

摘　　要：非采样Contourlet变换是一种新的多尺度多分辨率分析工具.本文提出了一种基于非采样Contourlet变换的彩色图像无监督分割算法.首先利用非采样Contourlet变换的平移不变性在其变换域应用梯度向量法提取图像多尺度边缘;然后在Contourlet变换域的低频子带和高频子带中分别提取局部低频能量纹理特征与高频多尺度Zernike矩纹理特征,并将二种纹理特征融合.最后在边缘图像中映射种子像素点,利用纹理和颜色特征欧氏距离,对彩色图像采用区域生长和区域合并的方法进行分割.实验结果证明:该算法将图像空间域的颜色特征与非采样Contourlet变换域的多尺度边缘和纹理特征恰当结合在一起实现彩色图像无监督自动分割,与传统算法相比有更高的准确性和鲁棒性.Nonsubsampled contourlet transform is a new multi-scale multi-resolution powerful analysis tool. An unsupervised segmentation algorithm for color image is proposed based on nonsubsampled contourlet transform. Firstly, for nonsubsampled contourlet transform shift invariance, multi-scale edge is extracted in transform domain by using gradient vector method. Then, local low-frequency energy texture features and high-frequency multi-scale Zernike moments texture features are extracted from low-frequency sub-band and high-frequency sub-band in transform domain and fusing them. Finally, detecting seed points in edge map to represent color image regions, the region growing followed by region merging method is applied for segmentation by color and texture Euclidean distance. The experimental results show that the algorithm can automatically fulfill unsupervised segmentation for color image by combining color, multiscale edge and texture properly, and has more precise and more robust segmentation effect than traditional algorithm.

关 键 词：图像分割 非采样CONTOURLET变换 多尺度边缘 纹理特征 区域生长 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]