基于成对约束Info-Kmeans聚类的图像索引方法  

Image indexing method based on clustering via Info-Kmeans under pair constraints

作　　者：刘文杰[1] 伍之昂[2] 曹杰[2] 潘金贵[1] 

机构地区：[1]南京大学软件新技术国家重点实验室,江苏南京210046 [2]南京财经大学江苏省电子商务重点实验室,江苏南京210003

出　　处：《通信学报》2013年第7期159-166,173,共9页Journal on Communications

基　　金：国家自然科学基金资助项目(71072172;61103229);江苏省省属高校自然科学研究重大基金资助项目(12KJA520001);国家科技支撑计划基金资助项目(2013BAH16F01);国家国际科技合作基金资助项目(2011DFA12910);江苏省自然科学基金资助项目(BK2010373;BK2012863)~~

摘　　要：针对图像数据噪声大和高维稀疏的特点,提出了一种基于噪声过滤和Info-Kmeans聚类的图像索引构建方法。首先,利用余弦兴趣模式过滤噪声。其次,提出了一种新的Info-Kmeans聚类算法,该算法不仅避免KL-divergence计算过程中的零值困境问题,还能融合以成对约束出现的先验知识。最后,在LFW和Oxford_5K 2个图像数据集上的实验表明:噪声过滤能显著提高聚类性能;Info-Kmeans比已有聚类工具具有更优越的性能。Constructing high-quality content-based image indexing is fairly difficult due to the large amount of noise in the data set and the high-dimension and the sparseness of the image data. To meet this challenge, a novel noise-filtering and clustering was proposed using Info-Kmeans based image indexing construction method. Firstly, a noise-filtering method using the cosine interesting patterns was presented. Secondly, a novel Info-Kmeans algorithm was proposed which could avoid the zero-feature dilemma caused by the use of KL-divergence and exploit the prior knowledge in the form of pair constraints. The experimental results on the two image data sets, LFW and Oxford 5K, well demonstrate that： noise filter can improve the clustering performance remarkably and the novel Info-Kmeans algorithm yields better results than the existing clustering tool.

关 键 词：图像索引 兴趣模式 噪声过滤 聚类分析 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]