一种改进的自然图像分割方法  

Improved Approach of Natural Image Segmentation

作　　者：李新征[1] 易建强[1] 赵冬斌 洪义平[1] 

机构地区：[1]中国科学院自动化研究所综合自动化工程中心

出　　处：《计算机应用研究》2006年第5期251-254,共4页Application Research of Computers

基　　金：科技部国际科技合作重点项目(2004DFB02100)

摘　　要：提出了一种改进的Mean Shift自然图像分割方法。首先从数据分析的角度出发,由数据本身估计用于聚类的颜色带宽。然后根据上层的视觉任务调整颜色带宽,调整后的颜色带宽可以看作是分割分辨率。根据调整后的颜色带宽,估计每个像素的密度值,并采用直接搜索的方法搜索局部密度极大值点。在局部模式检测之后。通过一个全局标准控制的局部模式融合,就可以将图像分割成最终的结果。在不同的光照条件和复杂的背景下,这个全局标准能达到稳定分割结果和控制最终区域数目的目的。This paper presents an improved mean shift method in segmentation of natural image. Firstly, the bandwidth of color, which is used for cluster, is estimated from image data. Then the bandwidth of color will be adjusted according to the upper vision tasks, and it can be regarded as the resolution of segmentation. The density value of every pixel is estimated according to the adjusted bandwidth of color, and the local maximum value of density is searched using direct density searching method. Lastly ,the final result of image segmentation can be got through a global optimization criterion for mode merging. Under different illumination conditions and complicated background, the global optimization criterion can achieve the purpose of stabilizing the result of segmentation and the number of the segmented regions.

关 键 词：带宽选择 图像分割 均值漂移 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]