A unified framework to support image database clustering and content-based image retrieval is proposed by the authors. For this purpose, they use the Markov model mediators (MMMs) mechanism to facilitate conceptual clustering. A MMM (triple M) is a stochastic finite state machine. Each state of a MMM is attached to a stochastic output process for describing the probability of occurrences of the output symbols (states). In the study, MMMs are used to explore affinity relations among the images at the database and cluster levels.

The authors consider both effectiveness and efficiency issues, and provide, respectively, some reasonable results and a short discussion. For the latter, they ignore the off-line cost of clustering and related activities, with the assumption that they will be performed annually or semi-annually. This sounds quite unrealistic in dynamic environments, and needs further attention. The authors seek to address the problems that involve large databases, and, in their experiments, they use 12 databases, with a total of 18,700 images. This is a good experimental number, but it leaves a lot of ground to cover in terms of scalability. In their discussion, the authors use the term “distributed databases,” and actually mean “autonomous databases,” that is, databases that involve no interrelated integrity constraints. They state what they mean by “distributed databases” in their text. However, for clarity, they should have used either the phrase “autonomous databases” or “multi databases,” instead of “distributed databases,” throughout the paper.

The material looks interesting, but it is hard to appreciate without additional reading. A nicely written, yet long general introduction could have been shortened, and the unused half-page at the end of this conference paper could have been used to provide more information about the authors’ related earlier work.