摘要
为了克服经典区域增长算法在复杂目标与背景分布情况下,停止条件难以确定的不足,提出基于目标模糊置信度描述驱动的区域能量进化增长图像分割算法.该算法结合了主动轮廓模型(Active contour model,ACM)、目标数据分布域描述与区域增长三者的优点,首先利用分割目标的支持向量数据域描述将待分割图像转化为相对于分割目标的模糊置信度表示,因为分割过程充分利用了有监督学习策略得到的目标特征分布情况,使得本文提出的算法具有更高的稳定性和更加广泛的适用范围,特别是对目标灰度分布不均或存在多纹理的目标也可以得到较好的分割结果.在区域增长进行分割时,引入了新的区域能量表示模型作为区域增长的结束判决条件,分割时逐渐降低目标模糊置信度的门限,通过对区域能量模型的动态优化来逼近最佳分割结果。对比实验结果表明本文提出的算法具有更大的灵活性和更好的分割性能。
To overcome the difficulty to search the stop condition in a conventional region growing algorithm, a novel region energy evolution image segmentation method is put forward, which couples the merits of support vector domain description, Mumford-Shah active contour energy model rand region growing. The input image data are transform into fuzzy object confidence description firstly by using the support vector domain description model, so the advantages of supervised kernel learning model and the global region distribution information could be exploited to enhance the segmentation performance. On the other hand, a new region-based image energy term in region evolution based on the fuzzy object confidence description is presented. It is more robust than the classical region growing rand active contour method, because it takes into account the optimal image object fuzzy confidence description knowledge of human being and feasible energy model as well. In the region growing processing step, the confidence threshold is updated gradually, so the optimal segmentation results are obtained by dynamic optimizing the novel energy model. Experimental results have demonstrated the flexibility and better performance of this novel region growing image segmentation method.
出处
《自动化学报》
EI
CSCD
北大核心
2008年第9期1047-1052,共6页
Acta Automatica Sinica
基金
北京大学视觉与听觉信息处理国家重点实验室开放基金(0507)
燕山大学博士基金(B287)
河北省自然科学基金(F2008000891)资助~~
关键词
图像分割
区域增长
支持向量数据域描述
模糊置信度
Image segmentation, region growing, support vector domain description （SVDD）, fuzzy confidence
