基于内容的半监督流形图像检索  

Manifold Based Semi-Supervised Learning for Content-Based Image Retrieval

作　　者：王志强[1] 傅向华[1] 赵良辉[2] 杜文峰[1] 

机构地区：[1]深圳大学计算机与软件学院 [2]五邑大学管理学院,江门市东城村22号529020

出　　处：《武汉大学学报（信息科学版）》2009年第8期928-931,935,共5页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金资助项目(60703112);广东省自然科学基金资助项目(7301329);深圳市科技计划资助项目(200741)

摘　　要：提出了一种半监督流形图像检索方法。给定查询图像,首先利用L1距离进行相似度计算,构造出图像数据库的加权图并排序;然后,由用户从排序结果中选择少量与查询图像相似的图像,形成正反馈图像集合作为已标注样本,其他图像作为未标注样本,进行基于拉普拉斯流形正则化的半监督学习,训练分类函数;最后利用训练好的分类函数对数据库中的图像进行分类排序,返回检索结果。在Corel图像数据集上的实验结果表明,半监督流形图像检索方法能获得较好的效果。We propose a manifold regularization based semi-supervised learning for image retrieval （MRBS2L） in this paper. Given query image, MRBS2L at firstly calculates the similarities between the query image and images in database with L1 distance and ranks, and constructs the weighted graph. Then it selects some similar images from ranked results by the user to form relevance feedback image set, which will be used to act as labeled samples and the remainder images are used as unlabelled data, and then trains classification function with laplaeian manifold regularization based semi-supervised learning. Finally, it classifies and ranks the images with trained classification function and returns retrieval results. The experimental results on Corel image database demonstrate the effectiveness of the proposed algorithm.

关 键 词：图像检索 流形学习 核学习 半监督学习 

分 类 号：P237.3[天文地球—摄影测量与遥感]