Iterative Refinement by Relevance Feedback in Content-Based Digi tal Image RetrievalM. E. J. Wood, N. W. Campbell, B. T. Thomas, Iterative Refinement by Relevance Feedback in Content-Based Digi tal Image Retrieval. ACM Multimedia 98. ISBN 1-58113-036-8, pp. 13–20. September 1998. PDF, 251 Kbytes.
Many image-database retrieval systems rely heavily on the success of one -shot queries, using optimised feature sets to obtain the best possible results. What is often missing from this approach is acceptance of the fact that the user knows considerably more about the query being made than can be conveyed in such relatively simple terms. If the query fails then the user must try and improve the description using only the available feature descriptors. This paper describes how a query system can exploit the user's knowledge to a higher extent by employing relevance feedback to iteratively refine queries at run-time. Subjects of interest are chosen by selection of regions from pre-processed, segmented images, giving access to object-specific, local information which is not possible in a global pattern-matching approach. After an initial retrieval attempt, feedback is given in the form of acceptance or rejection of images offered. This information is used as a collection of positive and negative training examples for a class-specific classification network by identifying clusterings in the data and the spread along feature axes. Each network consists of a set of Radial Basis Function nodes with a non-linear perceptron output layer. Network training is carried out off- line using the data gathered during an on-line query session with the user. The user can review and adjust the behaviour of the network in the next session. Over time, collections of these networks can be built into a hierarchical class database, resulting into highly useful retrieval tool specifically train ed for the nature of the user's database.