Lower-Level and Higher-Level Approaches to Content-Based Image Retrieval

This paper describes a content-based image retrieval system that employs both higher-level and lower-level vision methodologies separately and in conjunction for the retrieval of images containing large manmade objects. The goal is to use the lower-level analysis module to increase the capability of the higher-level analysis module, for queries where the structure exhibited by the manmade objects is important. Higher-level analysis is performed globally to extract structure by employing the elements of perceptual grouping to extract different shape representations for higher-level feature extraction from primitive image features. The shape representations include “L” junctions, “U” junctions and parallel groups. Lower-level analysis is performed globally by using Gabor filters to extract texture features. A manmade object region of interest extracted by using perceptual grouping is used as a frame for conducting lower-level analysis. Lower-level analysis may be performed without confinement to the region of interest, i.e., over the whole image. A channel energy model is utilized to extract lower-level feature vectors consisting of fractional energies in various spatial channels. The image database consists of monocular grayscale outdoor images taken from a ground-level camera.