Intelligent Systems Laboratory

Image Database Retrieval

In this project, we use two feature extraction methods and two decision methods that allow a user to input an image or a section of an image and to retrieve all images from a database having some section in them that is like the user input image.

The feature extraction methods used are the variances of gray level spatial dependencies computed from co-occurrence matrices, and the line-angle-ratio statistics constituted by a 2-D histogram of the angles between two lines and the ratio of mean gray values inside and outside the regions spanned by those angles. We also combine these two sets of features to make use of their different advantages.

A likelihood ratio is defined to measure the relevancy of two images one being the query image and one being a database image so that image pairs which had a high likelihood ratio were classified as relevant and the ones which had a lower likelihood ratio were classified as irrelevant. Also k-nearest neighbor rule is used to retrieve k images which have the closest feature vector to the feature vector of the query image in the high dimensional feature space.

To evaluate the performance of the algorithms we use a protocol that translates a frame throughout every image to automatically construct groundtruth image pairs for the relevance and irrelevance classes. As experimental databases we use the Fort Hood Dataset including 1,000 512x512 gray scale images for the RADIUS project, a remote sensing image dataset including 90 gray scale LANDSAT and other satellite images with different sizes, and finally the COREL Stock Photo Library 1 including 3,100 images from 31 categories.

Description last modified: May 14, 1998. More recent information can be found in the papers below.

Related Publications

Publications last modified: October 2000.