Feature Selection for Image Retrieval

K. Messer (University of Surrey, UK)

Abstract:

In this talk a method that facilitates an iconic query of an image/video database will be presented. A user specifies a query object on which a standard set of texture and/or colour properties are calculated. The same texture and colour attributes have been pre-computed and stored for all the images in the database. The unknown regions in the database images can then be classified by comparing the query feature data to the unlabelled feature data from the database.

Results on seismic and borehole datasets provided by {\it Shell Research} using two different types of classifier, a Gaussian based classifier and neural network will be shown which verify that the suggested method works well.

Knowing what features to calculate for a query is a problem. Using as many features as possible not only results in a computational costly system but can also give worse results then if a sub-set of those features had been used. Several traditional metho ds for carefully selecting the set of features will be compared which significantly reduce the feature set size without a corresponding degradation in performance.

A novel method of selecting inputs and hidden units for the the neural network will then be introduced. It is demonstrated that this algorithm out-performs these more traditional feature selection methods.