Skip to Main Content
We propose an object boundary descriptor that facilitates the use-features on and off an object boundary for image retrieval. A string is used to model multiple continuous image and shape feature values on an object boundary. On the basis of these feature values and their higher-order derivatives the Taylor expansion provides an approximation of feature values in the immediate neighborhood of the object boundary. This object boundary description is employed within an existing population-based incremental interactive visual concept learning method for image retrieval. A set of 245 vertebral X-ray images is used to measure effects off the proposed descriptor in terms of number of relevance feedback steps and precision versus recall. Results show increased efficiency and efficacy.
Date of Conference: 15-18 April 2004