摘要
由于逐对形状匹配不能很好地反映形状间相似度,因此需要引入后期处理步骤提升检索精度.为了得到上下文敏感的形状相似度,本文提出了一种基于期望首达时间(Mean first-passage time,MFPT)的形状距离学习方法.在利用标准形状匹配方法得到距离矩阵的基础上,建立离散时间马尔可夫链对形状流形结构进行分析.将形状样本视作状态,利用不同状态之间完成一次状态转移的平均时间步长,即期望首达时间,表示形状间的距离.期望首达时间能够结合测地距离发掘空间流形结构,并可以通过线性方程进行有效求解.分别对不同数据进行实验分析,本文所提出的方法在相同条件下能够达到更高的形状检索精度.
Since pairwise shape similarity analysis can not measure the shape distance accurately, post-processing steps are introduced into shape matching process for increasing retrieval scores. In this paper, a novel shape distance based on mean first-passage time （MFPT） is proposed for resolving the problem of learning context-sensitive similarity. Given the distance matrix computed by a distance function, discrete-time Markov chains are constructed for analyzing the underlying structure of the manifold formed by shapes. With each shape in the database regarded as a state, the mean number of steps for transition between two states, named mean first-passage time, is used to measure the shape distance. The mean first-passage time induced by geodesic paths can capture the shape manifold structure, and it can be obtained by solving the linear equations. Experimental results on different databases show that shape retrieval results can be effectively achieved by using the proposed method.
出处
《自动化学报》
EI
CSCD
北大核心
2014年第1期92-99,共8页
Acta Automatica Sinica
基金
国家自然科学基金(61374154
61074096)资助~~
关键词
形状匹配
形状距离学习
相似度矩阵
离散时间马尔可夫链
期望首达时间
Shape matching, learning shape distance, affinity matrix, discrete-time Markov chain, mean first-passagetime （MFPT）
