Image Retrieval Based on Structural Content
Sushil Bhattacharjee, Touradj Ebrahimi · 1999
A content-based image-retrieval system is described in this paper. The system is designed to administer a heterogeneous collection of images. The goal is to support querying in a rotation- and translation-, and juxtaposition-invariant fashion. The conceptual similarity measure used to compare two images is the number of small image-patches the images have in common. The patches to be compared are chosen using a 2D continuous wavelet which acts as a low-level corner detector. The local maxima in the response of this wavelet are used to locate potential corner-like features in the image. Then, a small region around every image-feature is used for the similarity measure. Each region is characterized by a set of Gaussian-derivative filter-responses evaluated at the corresponding feature-point. The filters extract local texture information from the luminance-channel of the image. The responses of $n$ such filters for one region are organized in a $n$-D vector, referred to as a token. These tokens are further quantized to select indexing-terms which are used to describe each image. Every image is represented by a vector of weights corresponding to an ordered set of indexing-terms. The task of comparing images then boils down to computing a vector product. The proposed IR system also supports query-refinement using the classical relevance feedback approach.