基于稀疏表示分类的图像检索方法.doc


基于稀疏表示分类的图像检索方法#
付海燕1,赫然2,孔祥维1,廖启1**
(1. 大连理工大学信息与通信工程学院,辽宁大连 116024;
5
10
15
20
25
30
35
40
2. 中国科学院自动化研究所模式识别国家重点实验室,北京 100190)
摘要:传统的图像检索基于底层图像特征的相似性,采用 K 邻近法进行检索排序。由于语
义鸿沟的存在,底层图像特征会影响检索结果的准确率,尤其当图像种类繁多时,检索的准
确率大幅下降。本文提出了一种基于稀疏表示的图像检索算法,稀疏表示是将底层特征从欧
氏空间转换到了稀疏空间,变成图像库字典集的线性组合,通过分析图像特征在稀疏空间中
的分布规律,在稀疏空间对图像困进行检索,可以有效的缓解语义鸿沟的问题。本算法在
Corel 数据库和微软商品图像库进行测试,实验结果表明,本算法的检索准确率和召回率优
于 K 邻近检索方法。特别是对于图像种类较多的图像库,检索效果更好。
关键词:图像检索;稀疏表示;稀疏空间;K 近邻检索
Image Retrieval Based Sparse Representation Classification
Fu Haiyan1, He Ran2, Kong Xiangwei1, Liao Qi1
(1. School of Information munication Engineering, Dalian University of Technology,
LiaoNing DaLian 116024;
2. Chinese Academy of Sciences Institute of automation of National Laboratory of pattern
recognition, Beijing 100190)
Abstract: K-nearest neighbors searching is used in traditional image retrieval, which ranks images
based on their low-level feature similarity. However, the choice of low-level features will produce
an effect on retrieval precision due to the semantic gap, even when image categories increases. An
image retrieval method based on sparse representation is proposed in this paper. Low-level
features are transferred to sparse space from Euler space, and are represented as a linear
combination of train images, which narrows the semantic gap. Then, we retrieve images by
analyzing the distributio