摘要
针对经典区域增长算法中生长规则以及特征选取困难的问题,提出基于可拒识双层支持向量机模型的多目标并行区域增长图像分割算法。首先交互选择多个不同区域的种子点,并交互选择属于每个目标区域的子块和非目标区域的子块形成双层支持向量机训练样本;然后利用这些已知的训练样本训练双层支持向量分类器;在区域增长过程中,首先利用第一层的最大间隔超平面的支持向量分类器(SVC)进行分类判决,对属于该区域的点再利用第二层的支持向量域数据选择器(SVDD)进行拒识或接受处理,最后利用两层分类器的结果进行综合判决。为避免初始种子点位置选择对算法性能的影响,采用了多区域并行竞争增长策略。仿真实验获得了较好的分割效果,实验结果表明,提出的算法是合理可行的。
A multiple-object parallel growing algorithm based on two-layer support vector machines with rejection feature is proposed to solve the difficulty in feature selection and region growing rule in the conventional region growing image segmentation algorithm. Initial seeds in each object are selected by interactive manual operation. At the same time the blocks that belong to or don＇ t belong to the object region are interactively selected. Then the two-layer support vector classifier is trained by the training data collected in the first step. In region growing process, firstly the first layer of support vector classifier （SVC） with maximum margin between two classes is used for classifying the input block; then the sphere support vectors of this object region to describe the distribution of the sample are obtained by searching all the sphere boundaries containing the samples of this region. Then the input pattern of no-object classes could be rejected by the second support vector domain description （SVDD）. In order to make results independent of the processing order and the initial growing seeds, multiple-object parallel growing is employed. Experimental results show that this algorithm is feasible and it performs better than conventional region growing algorithm.
出处
《系统工程与电子技术》
EI
CSCD
北大核心
2006年第5期677-680,共4页
Systems Engineering and Electronics
基金
河北省科学技术研究与发展项目(Z2005310)
2005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)资助课题
关键词
图像分割
区域增长
支持向量机
image segmentation
region growing
support vector machine
