基于对象的自然图像层次分割  

Object-based Hierarchical Segmentation of Natural Image

作　　者：徐涛[1] 张立民[2] 张建廷[1] 

机构地区：[1]海军航空工程学院电子信息工程系,烟台264001 [2]海军航空工程学院科研部,烟台264001

出　　处：《系统仿真学报》2013年第9期2050-2054,共5页Journal of System Simulation

基　　金：国家自然科学基金(61102167)

摘　　要：基于像素的图像分割方法在表示同质区域的像素之间关系上存在局限性,其实现算法往往具有较高的计算复杂度。针对机器视觉应用中高质量自然图像分割算法运算复杂度过高的问题,构建了一种基于对象的自然图像层次分割架构。采用多尺度归一化割对图像进行预分割,同分水岭变换及Voronoi图法相比,该方法可以采用很少的图像对象就能较好的保持图像结构,且运算时间与图像尺寸和所需的图像预分割区域数成线性关系。提出了一种改进的超度量相异性定义,将图像对象作为整体进行处理,更为充分的利用了图像对象的区域属性。实验证明方法在获得较好的分割效果的同时,具有很高的运算效率。Pixel-based image segmentation algorithms have limitation on representation of long range interactions between pixels belonging to the same homogeneous region, and usually are computationally inefficient. A framework of natural image hierarchical segmentation with methodology of object-based image analysis was proposed. The framework aimed to reply the inefficiency of high quality segmentation algorithm used in machine vision. The input image was pre-segmented by multiscale Normalized cuts. This method could maintain the geometric structure of the image with less segments comparing to Voronoi tessellation and watershed transform, and could be computed in linear time. A new ultrametric dissimilarity of image objects was defined. This measure of dissimilarity took sufficient advantage of the property of the image object. Experiments show that the framework works efficiently with high quality segmentation result.

关 键 词：自然图像分割 基于对象图像分析 层次分割 超度量相异性 索引层次结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]