基于多核学习和AP聚类的图像摘要选取方法

Image summarizing based on multi-kernel learning and AP clustering

针对交互式图像检索系统在特征空间中选取摘要图像时,遇到多种特征难以融合及聚类效果不佳等问题,提出了一种基于多核学习和AP(affinity propagation)聚类的图像摘要选取方法。先利用多核函数融合多种特征计算图像间的相似度,再进行AP聚类,并选取聚类中心图像作为整个图像集的摘要。实验表明,所选摘要图像能直观有效地反映原图像集的内容梗概,据此对原图像集的查询拥有较高的查全率和查准率。

Image summarizing is a critical step in the interactive image retrieval.The traditional methods taking in the feature space may encounter with the feature fusion and bad clustering results.This paper developed a novel approach based on multi-kernel learning and AP clustering.Firstly,got the image similarity using multi-kernel function to fuse the diverse visual features.Secondly,clustered the images into the compact classes with AP method.Lastly,selected the exemplars as the summarization for the whole image collection.The experiments show that the image summarization selected by the method above can give a good global overview of image collections and performance of retrieval.