When designing interfaces for image databases, current database ideology suggests query-based interfaces expressed in SQL or similar languages (e.g. Cardenas et al, 1993) We believe this approach is not well suited to typical use. Query-based interfaces are hard for most non-experts to use, and wearying even for experts. Users must possess knowledge about the domain of the database, the functionality of the retrieval system, and the classification scheme employed by the database; even experienced users may have difficulty constructing effective queries (Chen & Dhar, 1990). Also, the classification task, since done by human beings, takes enormous time and other resources as the database grow large.
Further, the interface programmer has to deal with complicated interface design or language issues. The generic question continually running through the minds of most interface programmers often seems to be: What should I have the interface do if the user makes this particular mistake? Rather than focussing on the user's errors, we wish to focus on the user's actions and leave more of the classification task up to the system.
This paper develops an alternative with some desirable attributes:
A prototype of this system is currently working on a database of over 10,000 images with a 20-processor Silicon Graphics machine and a 2-gigabyte disk.