基于密度聚类和多示例学习的图像分类方法  

Image categorization method using density clustering on region features and multi-instance learning

作　　者：陈涛[1] 邓辉舫[1] 刘靖[1] 

机构地区：[1]华南理工大学计算机科学与工程学院,广州510006

出　　处：《吉林大学学报（工学版）》2014年第4期1126-1134,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(61003174)

摘　　要：针对图像的低级特征表示与高级概念之间的语义鸿沟,本文利用密度聚类获得的簇分布信息和多示例学习框架在区分歧义性对象上的特点,提出了一个基于区域特征密度聚类和多示例学习的图像分类方法(DCRF-MIL)。该方法首先将每个图像分割为多个区域,将所有区域组成一个集合,在这个区域集合上,使用密度聚类算法学习到区域特征的簇分布信息;然后,将图像看作包,区域看作包中的示例,基于区域特征的簇分布信息,将包映射为簇分布空间上的一个向量作为包的特征,使得包特征带有图像区域的语义信息;最后,使用支持向量机算法,在带有包特征的训练集上训练分类器,对测试图像进行分类。在Corel图像集和MUSK分子活性预测数据集上的实验表明,DCRF-MIL算法具有分类精度高和参数易于选择等特点。In order to narrow the semantic gap between low-level visual features and high-level semantic concepts in image categorization, the clustering information from a density clustering algorithm and the characteristics of multi-instance learning framework in distinguishing ambiguous object are exploited. An image categorization method is proposed using Density Clustering on Region Features and Multi-Instance Learning （DCRF-MIL）. First, the DCRF-MIL divides each image into number of regions and relines up all regions into a collection; then it uses a density clustering algorithm to learn the potential distribution information of the region features in the collection. Second, it treats an image as a bag and the regions as instances in the bag. Based on the cluster distribution information of region features, the bag is mapped into a vector in the cluster distribution space. Finally, a support vector machine classifier is constructed to predict the class label of the unlabeled image. The experiments on the Corel image data set and MUSK molecular activity prediction data set show that the DCRF-MIL method has high classification accuracy and it is easy to select its parameters.

关 键 词：人工智能 图像分类 多示例学习 密度聚类 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]