基于自适应尺度LE的轮廓编组算法

在图嵌入框架下，以编组线索的聚类来实现轮廓编组的方法，不仅可以有效地将局部特征和全局特征结合起来，而且更加符合人类视觉感知以流形存在的特点．本文在分析相似度矩阵对样本结构表示意义的基础上，提出了一种基于自适应尺度LaplacianEigenmap的轮廓编组算法．该算法能够根据编组线索多维特征的不同局部统计特性，自适应地改变相似度计算模型中的尺度因子，使相似度矩阵更准确地反映编组线索数据集的结构特性．在此基础上通过降维实现编组元的聚类，从而得到轮廓编组的结果．实验证明，该算法对于局部统计特性差异较大的编组线索有着良好的适应性，尤其对于有遮挡的感知目标，表现出比图分割方法更为优越的性能．

Contour grouping algorithm based on Laplacian Eigenmap with adaptive scale

In this paper, contour grouping is studied under the general framework of graph embedding by clustering the grouping cues. In perceptual grouping process, local features and global features are combined and manifold learning is used to obtain grouping result by simulating the human vision. By analyzing the meaning of affinity matrix for representing the structure of data, a contour grouping algorithm based on the Laplacian Eigenmap with adaptive scale is proposed. According to the different local statistical features, the scale factor is automatically determined to compute the affinity between each pair of group-meta which represent the structure feature of grouping cues more exactly. As a result, contour grouping is implemented by clustering the group-meta cues based on the result of dimen- sionality reduction. The experiments show that this algorithm has the good adaptability for local statistical discrepancy and shows the higher efficiency than the normal method of graph segment, especially for partially occluded object.