大规模图像集中的代表性图像选取

Representative Image Selection from Image Dataset

针对传统图像检索系统通过关键字搜索图像时缺乏语义主题多样性的问题,提出了一种基于互近邻一致性和近邻传播的代表性图像选取算法,为每个查询选取与其相关的不同语义主题的图像集合.该算法利用互近邻一致性调整图像间的相似度,再进行近邻传播(AP)聚类将图像集分为若干簇,最后通过簇排序选取代表性图像簇并从中选取中心图像为代表性图像.实验表明,本文方法的性能超过基于K-means的方法和基于Greedy K-means的方法,所选图像能直观有效地概括源图像集的内容,并且在语义上多样化.

In a traditional image retrieval system, people search images using keywords. However, the result shows a lack of diversity in the sense of semantic theme. For the problem, we propose a viable method for representative image selection. We define representative images as those with diverse contents in the semantic meaning to cover different semantic forms of a query. First, we use mutual nearest neighbor consistency to adjust the similarity between images as the input to the AP clustering. Then we select representative clusters based on cluster ranking and finally take the images of the cluster center from representative clusters as a summary of the image dataset. The results showed that the performance of our method is better than the K-means based method and the greedy K-means based method. The selected images can summarize the content of the original image dataset intuitively and effectively, and they are diverse in semantic meaning as well.