一种基于广义期望首达时间的形状距离学习算法

发布时间：2018-05-22 20:55

本文选题：形状匹配 + 形状距离学习　；参考：《自动化学报》2016年02期

【摘要】：形状距离学习是形状匹配框架中引入的后处理步骤,能够有效改善逐对计算得到的形状间距离.利用期望首达时间分析形状间相似度可能导致距离更新不准确,针对这一问题提出了一种基于广义期望首达时间(Generalized mean firstpassage time,GMFPT)的形状距离学习方法.将形状样本集合视作状态空间,广义期望首达时间表示质点由一个状态转移至指定状态集合所需的平均时间步长,本文将其视作更新后的形状间距离.通过引入广义期望首达时间,形状距离学习方法能够有效地分析上下文相关的形状相似度,显式地挖掘样本空间流形中的最短路径,并消除冗余上下文形状信息的影响.将所提出的方法应用到不同形状数据集中进行仿真实验,本文方法比其他方法能够得到更准确的形状检索结果.
[Abstract]:Shape distance learning is a post-processing step introduced in shape matching framework, which can effectively improve the distance between shapes calculated by pair. Using expected first arrival time to analyze shape similarity may lead to inaccurate distance updating. To solve this problem, a shape distance learning method based on generalized expected first arrival time and generalized mean firstpassage time GMFPTs is proposed. The shape sample set is regarded as a state space, and the generalized expected first arrival time represents the average time step required for a particle to be transferred from a state to a specified state set. In this paper, it is regarded as the distance between shapes after updating. By introducing generalized expected first arrival time, shape distance learning can effectively analyze context-dependent shape similarity, explicitly mine the shortest path in the sample space manifold, and eliminate the influence of redundant context shape information. The proposed method is applied to different shape datasets for simulation experiments. Compared with other methods, the proposed method can obtain more accurate shape retrieval results.
【作者单位】：大连理工大学电子信息与电气工程学部;
【基金】：国家自然科学基金(61374154) 中央高校基本科研业务费专项资金(DUT14RC(3)128)资助~~
【分类号】：TP181

【参考文献】