In this paper, we discuss a number of new problems that arise in image databases, and that set them apart from traditional databases. The fact that image databases are based on similarity, rather than matching, creates a whose set of new issues.Most noticeably, while matching is, by and large, a well defined concept, there are many possible types of similarities. In this paper, we consider the problem of simulating human similarity perception. We argue that a satisfactory solution is possible for preattentive similarity, and we present a general and comprehensive geometric similarity model.