嵌入标记信息的铁路扣件状态检测主题模型  

Topic Model of Railway Fastener State Detection Embedded with Tag Information

作　　者：欧阳[1] 罗建桥[1] 李柏林[1] 李爽 OU Yang,LUO Jianqiao,LI Bailin,LI Shuang(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,Chin)

机构地区：[1]西南交通大学机械工程学院,成都610031

出　　处：《计算机工程》2018年第6期200-206,共7页Computer Engineering

基　　金：四川省科技支撑计划项目(2016GZ0194)

摘　　要：针对潜在狄利克雷分布(LDA)模型忽略特征单词明确性的问题,提出一种嵌入标记信息的主题模型WL_LDA。设计一种基于SIFT特征点约束单方向LBP图像的方法。运用该方法获取图像的纹理结构,对视觉单词进行标记。将标记信息嵌入到LDA中,利用单词和标记的二维直方图推导图像的主题分布。通过运用该主题分布训练分类器,完成铁路扣件的状态检测。实验结果表明,与LDA主题模型相比,各扣件在主题空间中的区分度增加4.5%~15%,与现有PCA、DF等方法相比,漏检率和误检率明显降低,具有较好的分类性能。Aiming at the explicit problem of ignoring the characteristic words in the Latent Dirichlet Allocation （LDA） model,a topic model WL_LDA embedded tag information is proposed.A method for constraining unidirectional LBP images based on SIFT feature points is designed to obtain the texture structure of the image and the visual word is marked.By embedding the tag information in the LDA,the topic distribution of the image is derived using the two-dimensional histograms of words and tags.This topic distribution is used to train the classifier to complete the state detection of railway fasteners.Experimental results show that,compared with the LDA topic model,the differentiation degree of various types of fasteners in the topic space increases by 4.5% to 15%.Compared with the existing methods such as PCA and DF,the rate of missing detection and false detection is significantly reduced,it has better classification performance.

关 键 词：图像语义分析 潜在狄利克雷分布 视觉单词 SIFT特征 单词标记 主题模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]