Hierarchical indexing for region-based image retrieval

Document Type

Conference Proceeding

Publication Date

12-1-2004

Abstract

A region-based image retrieval system applies image segmentation to divide an image into discrete regions, which if the segmentation is ideal, correspond to objects. Image retrieval can be very effective since each image in the database is described by features that illustrate the object semantically. The focus of this research was to improve upon the capture of regions so as to enhance indexing and retrieval performance. For new images, image pixels are first segmented into different regions based on color, texture and shape features. Images are first partitioned into 4 by 4 blocks to compromise between texture granularity and computation time. On each block, six features are used for segmentation. Three of them are the average color components in the perceptually uniform LUV color space, where L encodes luminance, and U and V encode color information. The other three are texture features represented by Daubechies-4 wavelet transform [1]. A modified version of the k-means algorithm [2] is then used to cluster the pixel feature vectors into region clusters. The k-means algorithm is useful as it does not require prior knowledge about the number of regions present. Each region is saved to the image database and treated as a single image for indexing purposes. On each region, shape features are then defined as a vector containing three components of normalized inertia [4] of order 1 to 3 of the region. For indexing, a second level of clustering is performed. A rotationally invariant Euclidean L2 distance metric is applied to the color vector, texture vector and shape vector respectively to obtain overall similarity between two images. The metric is then used to cluster regions into classes. To make the image query faster and scalable, a hierarchical clustering algorithm using a tree structure vector is introduced which saves on storage as well as facilitates retrieval efficiency. At search time, an example query image (or set of images) is presented. The query image is initially segmented into regions by the methods noted above, and then compared with the hierarchical cluster centers. Once the closest clusters are identified, the regions are then compared to the subcluster centers to further refine the search. Once the lowest cluster level is reached, all images in the matched clusters are returned to the user. As part of this process, the importance of a region within an image is given consideration by giving more weight to foreground objects than to background, and to objects located closer to the center than to those whose centroid is closer to the edges. A prototype system has been developed to implement this process. The system is being tested against some well known region-based retrieval algorithms such as IRM [4] and FuzzyClub [5] methodologies in order to compare performance. Preliminary testing has been promising. Results of these tests will be presented at the IERC conference.

Publication Source (Journal or Book title)

IIE Annual Conference and Exhibition 2004

First Page

2231

This document is currently not available here.

Share

COinS