sets and applications, e.g., [4][5][6]. General-purpose image database management systems either define visual rele-Operational image databases are subsystems of systems which contain images, structured data, structured data complevance in terms of low level features and give the user tions, and algorithms, all relating to a particular real-world control over how features are used in the distance measure, domain. Traditional notions of image similarity therefore need e.g., [7], or provide comprehensive suites of relevance meato be extended to handle retrieval of mixed datatypes. In this sures each suitable for specific application areas, possibly paper, we use the structured datatypes to describe a user's with some self-learning capability [8]. In any case, different interests and background knowledge and then to ascribe a value users have different notions of relevance, so the option of to that user of the information in the system. The difference weighting components of the measure is provided, e.g., between the information values in two states of a system leads [9, 10].
naturally to a user-dependent definition of the (dis) similarity
It is obvious that the more knowledge there is available between the states. These definitions restrict respectively to to the system about the domain in which the user is interdefinitions of the value of information in an image, and of the ested, the more likely it is that the system can match the (dis)similarity between two images. We show how this notion user's notion of relevance. In most operational situations, of similarity provides an integrated way of dealing with image and nonimage data, generalizing conventional classes of feaa user who retrieves an image is accessing an information ture-based similarity measures and allowing for missing data. system which also contains non-image data and algorithms, Because the measure depends on the data models and data gathered together to support the activities in which that in the structured part of the database as well as on image user is engaged, and so providing information about the characteristics, the similarity between two images may change domain of interest. Relatively mature examples of such when nonimage data in the system are updated or extended. systems are found in geographic, medical, and desktop We illustrate using an example from a business domain. ยฉ 1997 publishing application areas, and will become increasingly
Academic Press
common as commercial organizations start to rely on multimedia data [11]. These systems will allow nonimage data to be used to tune image retrieval mechanisms. But over 97