一种结合PCA与信息熵的SIFT特征向量自适应降维算法  

Self-adaptive Descending Dimension Algorithm of SIFT Feature Vector Combined with PCA and Information Entropy

作　　者：宋丹[1,2] 石勇[3] 邓宸伟[4] 

机构地区：[1]中国工商银行博士后科研工作站,北京100032 [2]中国科学院大学博士后科研流动站,北京100190 [3]中国科学院大学管理学院,北京100190 [4]北京理工大学信息与电子学院,北京100081

出　　处：《小型微型计算机系统》2017年第7期1636-1641,共6页Journal of Chinese Computer Systems

基　　金：国家"八六三"高技术研究发展计划项目(2015AA8017032)资助;国家自然科学基金项目(61301090)资助

摘　　要：针对尺度不变特征变换(SIFT)描述子在目标匹配与识别过程中特征向量维数较大、匹配实时性较差的问题,提出一种结合主成分分析(PCA)与图像信息熵的SIFT特征向量自适应降维算法.首先分析SIFT特征向量矩阵的稀疏性,用170组不同场景图像提取的SIFT特征向量作为训练数据,并对其进行PCA投影,生成特征向量PCA矩阵组;其次根据PCA矩阵能量层次与图像信息熵的关系筛选出PCA模板并拟合出"维度-信息熵"关系式;最后,加载PCA模板及"维度-信息熵"关系式,根据测试图像的信息熵,构建自适应降维算法,对测试图像的SIFT特征向量进行降维.本文采用多组图像验证本算法的匹配性能,实验结果表明,本算法能够针对不同场景进行自适应降维,降维后特征点匹配过程耗时降低,匹配精度与降维前相近.A novel self-adaptive descending dimension algorithm of SIFT feature vector based on PCA and image information entropy is proposed to solve the large numbers of dimension of SIPT feature vector and poor real-time capacity in the process of object matching and recognition. Firstly, this paper analyzes the sparsity of SIFT feature vector and uses 170 groups of images with different scenes as training data to extract SIFT vector. After that,using PCA to project with SIFT vector and a group of PCA matrices is generated. Secondly, this paper selects PCA module and ＂dimension-entropy＂ formula. Lastly, loading PCA module and ＂dimension-entropy formula＂, then using entropy of test images to build adaptive descending dimension algorithm. In this paper, we use many groups of images to test the performance of this algorithm. The result shows that this algorithm has the ability to descend dimension self-adaptively in different scenes and the real-time has improved after descending dimension, the matching accuracy is the same as SIFT with- out descending dimension.

关 键 词：SIFT描述子 主成分分析 信息熵 特征降维 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]