协同视觉信息与标注信息图像聚类

针对如何有效地利用图像视觉信息与标注信息进行图像聚类的问题,提出了一种基于视觉单词与标注单词共生的聚类算法.在视觉特征空间,采用K-means算法对图像聚类,得到表征图像视觉信息的视觉单词,即聚类中心.在图像标注字空间,计算各聚类中心下标注单词的统计分布,建立视觉单词与标注单词共生矩阵,进而针对图像提取嵌入有视觉信息的标注词特征LDA(latent dirichlet allocation)主题模型作为最终聚类算法完成图像的聚类.通过对Pascal VOC 2007标注图像数据库进行的实验仿真以及对比试验结果表明,基于视觉单词与标注单词共生的聚类算法可以有效地利用图像的视觉信息与标注信息的互补特性,提高聚类算法的性能.

Image Clustering via Combined Visual and Annotation Information

To effectively employ annotation information and visual information for image clustering,a novel algorithm based on the co-occurrence of the visual and annotation words is proposed.In the vision feature space,Kmeans algorithm is utilized to cluster the feature into visual words,namely the cluster centers.In the annotation word space,a visual-annotation co-occurrence matrix is constructed by computing the statistical distribution of the annotation words under each corresponding visual cluster.Then,the annotation feature with its corresponding visual information embedded can be extracted.Finally the LDA （latent dirichlet allocation） topic model is used to cluster the images.The numerical experiments on Pascal VOC 2007 database show that the proposed method can effectively take advantage of the complementary visual and annotation information to improve the performance of clustering algorithms.