基于内容图像检索中的优化鉴别特征  

Optimized Discriminative Features for CBIR

作　　者：施智平[1] 李清勇[2] 赵晓东 何清[4] 史忠植[4] 

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]北京交通大学计算机与信息技术学院,北京100044 [3]内蒙古自治区人民检察院技术处,呼和浩特010050 [4]中国科学院计算技术研究所智能信息处理重点实验室,北京100190

出　　处：《计算机辅助设计与图形学学报》2012年第12期1592-1598,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60933004;61035003;60903141;61072085);国家"九七三"重点基础研究发展计划项目(2007CB311004)

摘　　要：为了提高基于内容图像检索系统的检索速度和准确率,提出一种融合两类线性鉴别分析的方法来提取低维的优化鉴别特征.首先把多类问题转换为多个两类问题,对每个两类问题进行线性鉴别分析,得到鉴别向量;所有的鉴别向量组成鉴别变换矩阵,对图像特征进行投影变换得到鉴别特征;最后用变换后的鉴别特征进行图像检索或分类,得到准确率更高的结果.该方法中鉴别特征空间的维数与类别数相等.与多种特征优化方法进行比较的实验结果表明,采用文中方法可以显著地提高图像检索和图像分类的性能.In this paper, a method merging 2-class linear discriminant analysis is proposed to capture low-dimensional optimal discriminative features to improve the searching speed and precision of content-base image retrieval systems. First, a multi-class problem is translated to multiple 2-class problems with linear discriminant analysis to estimate a discriminant vector for each. Second, all the discriminant vectors are merged into a discriminant transformation matrix, by which image visual features are transformed into discriminant features. Finally, the discriminant features are employed to gain high precision of image retrieval and classification. The dimensionality of the discriminant features corresponds to the number of classes involved. The experiments, in which our proposed method is compared with various feature optimizing methods, show that the proposed approach improves the performance of image retrieval and classification dramatically.

关 键 词：两类线性鉴别分析 图像检索 图像分类 鉴别特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]