CMVF: A Novel Dimension Reduction Scheme for Efficient Indexing In A Large Image Database

• Main Authors: SHEN, Jialie, Ngu, AHH, Shepherd, John, Huynh, Du Q.
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2003
• Subjects: Databases and Information Systems; Numerical Analysis and Scientific Computing
• Online Access: https://ink.library.smu.edu.sg/sis_research/1239
• Institution: Singapore Management University

In recent years, due to the increasing volumes of multimedia data in the WorldWideWeb, digital library, biomedicine and other applications, efficient content based similarity search in large image databases is gaining considerable research attentions. As a result, various indexing methods known as Spatial Access Methods (SAMs) and metric trees have been proposed to support this kind of retrieval. The former includes SStree, R+tree and grid files; the latter includes the vptree, mvptree, GNAT and Mtree[3]. However, the optimised distance-based access methods currently available for multidimensional indexing in multimedia databases are developed based on two major assumptions: a suitable distance function is known a priori and the dimensionality of the image features is low. Unfortunately, these assumptions do not make the problem substantially easier to solve. For example, it is extremely difficult to define a distance function that accurately mimics human visual perception in image similarity measurement. Also, typical image feature vectors are highdimensional (dozens of dimensions). The standard approach to reducing the size of feature vectors is Principle Component Analysis (PCA). However, this approach might not always be possible due to the nonlinear correlations in the feature vectors. Motivated by these concerns, we proposed and developed the CMVF (Combining MultiVisual Features) framework, a fast and robust hybrid method for nonlinear dimensions reduction of composite image features for indexing in large image database[2]. This method incorporates both the PCA and nonlinear neural network techniques to reduce the dimensions of feature vectors, so that an optimised access method can be applied. In this demonstration, we show that with CMVF approach a small but welldiscriminating feature vector can be obtained for effective indexing. It allows us to incorporate classification information based on human visual perception into the indexing. In addition, effectiveness of the indexing can be improved significantly with integration of additional image features. In the following sections, we overview the design and system architecture of our CMVF system, and give performance evaluation.