Information retrieval using local linear PCA
Abstract
Efficient and accurate information retrieval (IR) is one of the main issues in multimedia databases. Clustering can help to generate the efficient indexing structures and provide the comparison between data types. The Most Expressive Feature (MEF) extraction can improve comparison accuracy between two data which belong to a same data type since it discards redundant features. The authors introduce a local linear principal component analysis (LLPCA) to design an optimal scheme for IR. The LLPCA realizes the clustering and local MEF extraction at the same time. Using these clusters and local MEFs, an IR scheme can be divided into two steps from coarse to fine. We apply the scheme to a trademark retrieval system to evaluate its performance based on the accuracy and efficiency measurements. The experimental results indicate this retrieval scheme is superior the other schemes using the original features or global MEFs extracted by a Global Linear PCA (GLPCA).
Authors: Xiaoqing Li, Irwin King