LDA-CRF:一种基于概率图模型的目标检测方法  

LDA-CRF:Object Detection Based on Graphical Model

作　　者：郭乔进[1,2] 李宁[1,2] 杨育彬[1,2] 武港山[1,2] 

机构地区：[1]南京大学计算机科学与技术系,南京210093 [2]计算机软件新技术国家重点实验室(南京大学),南京210093

出　　处：《计算机研究与发展》2012年第11期2296-2304,共9页Journal of Computer Research and Development

基　　金：国家"九七三"重点基础研究发展计划基金项目(2010CB327903);国家自然科学基金项目(60875011;60723003;61035003;60975043);江苏省自然科学基金重点项目(BK2010054)

摘　　要：目标检测和识别是计算机视觉和机器学习的研究热点.近年来,主题模型(如LDA等)在无监督的图像识别和定位等应用中获得了巨大的成功.然而,LDA忽略了图像区域之间的空间关系,并且不能处理连续值的视觉特征或特征向量.而条件随机场(CRF)能够利用图像区域之间的局部相关性来提高分类准确性.基于LDA和CRF提出了一种LDA-CRF模型.通过利用LDA生成的主题信息来辅助CRF的分类,同时结合图像区域之间结构化的类别信息来改进LDA的主题生成机制.实验结果表明,LDA-CRF模型的检测效果要优于CRF.Object detection and recognition is actively studied in computer vision and machine learning. Particularly, in recently years, topic models such as latent Dirichlet allocation （LDA） has achieved great success in unsupervised recognition and localization of objects. However, LDA ignores the spatial relationships among image regions. To address this issue, conditional random field （CRF） introduces local dependence to improve the classification accuracy of image patches. In this paper, we propose a latent Dirichlet allocation-conditional random field （LDA-CRF） model by combining LDA with CRF. CRF is trained with topic features generated by LDA, while LDA generates topic information by utilizing structured class labels provided by CRF. Experimental results show that LDA CRF performs better than CRF in object detection and recognition.

关 键 词：LDA CRF 目标检测 变分推论 概率图模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]