Interactive Searching of Image Databases
This project aims to develop a computer vision system which is capable searching digital image databases for user-defined objects. Databases of interest would include digital home photograph albums and also such databases as are found on the World Wide Web. Such databases can often be too large to allow every image to be inspected visually, and automatic techniques are required to assist with searching and interpretation.An untrained user defines objects of interest using a graphical interface and similar objects in the database are then found automatically. Refinement of the search is performed as the user identifies further positive and negative examples of the object of interest.
One of the fundamental problems in pattern recognition systems of this
kind is to determine a set of features appropriate for a specific
recognition task from the large set of features
that could be
potentially useful. For a general purpose system one cannot rely on
the user to make a suitable choice and this selection must, therefore,
be made automatically. However, to exhaustively seek for useful
subsets among the set of all possible features would lead to a
computational explosion.
Radial basis functions are used to model the data rather than the existing approaches using multi-layer perceptrons (MLP's). Radial basis functions can be trained in a similar manner to MLP's but have a simpler mathematical interpretation that allows the initial parameters to be determined using clustering algorithms. This allows us to perform preliminary classification of the data before sufficient examples are available to perform full optimisation. Such an approach permits us to extend the network and refine the feature set as more examples are identified. More information is available on http://www.cs.bris.ac.uk/Research/Vision/search.html.