基于Dirichlet过程混合模型的指节偏移测度特征学习  

Knuckle offset measure feature learning based on infinite Dirichlet process mixture model

作　　者：杨世强[1] 乔丹 弓逯琦 李德信[1] YANG Shiqiang;QIAO Dan;GONG Luqi;LI Dexin(School of Mechanical and Precision Instrument Engineering,Xi an University of Technology,Xi an 710048,China)

机构地区：[1]西安理工大学机械与精密仪器工程学院

出　　处：《计算机集成制造系统》2019年第9期2305-2313,共9页Computer Integrated Manufacturing Systems

基　　金：国家自然科学基金资助项目（51475365）;陕西省自然科学基础研究计划资助项目（2017JM5088）~~

摘　　要：针对人机协调装配光强相对稳定的室内环境,为获取指节图像特征,丰富手部位姿信息,提出基于无穷Dirichlet过程混合模型的指节识别方法。在局部Markov假设的基础上,将条件随机测度的学习问题转化为随机聚类学习问题;运用无穷Dirichlet过程混合模型描述概率密度,将聚类数量表述为随机态;利用Gibbs采样方法,对分层概率形式的密度结构进行迭代学习;给出了基于Dirichlet过程混合模型的坍塌Gibbs采样算法,运用样本集进行了模型训练学习。最后,采用固定阈值对手部图像指节进行识别,验证了所提方法的有效性。In indoor environment with relatively stable light intensity for man-machine coordinated assembly,the finger image feature should be extracted and recognized to accurately present the hand posture.A knuckle recognition method was presented to enrich the hand information of co-operator which based on infinite Dirichlet process mixture model.Based on the local Markov hypothesis,the learning problem of conditional randomness measure was transformed into a stochastic clustering learning problem.An infinite Dirichlet process mixture model was used to construct the probability density,and the number of clusters was described as a random state.The Gibbs sampling algorithm was used to iteratively learning the density structure of hierarchical probability.The collapse Gibbs sampling algorithm based on Dirichlet process mixed model was established and the model training was carried out with the sample sets.The knuckle image recognition was carried out with the fixed threshold and the effectiveness was validated.

关 键 词：无穷Dirichlet过程 GIBBS采样 图像识别 指节偏移测度 随机聚类学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]