Predicates and superimposed coding: An application to image recognition

Abstract

First of all we show that the evaluation of predicates leads to a classical problem of retrieval on secondary keys. After a brief sketch of the signature file technique we collect some known probabilistic results. This allows us to treat two distinct mathematical models simultaneously. We then consider the so-called false drop probability in the context of key-based image recognition. It is shown that known optimality results do not necessarily remain valid in this situation. This suggests shifting the emphasis from optimizing the false drop probability to optimizing the search time (both aims do not necessarily coincide). Thus we are led to consider a 2-level scheme which is simple compared to other 2-level schemes which have been used before. This scheme can be employed because of a suboptimal choice of parameters. Finally theoretical predictions derived from a somewhat crude abstract model are validated by experimental work carried out on a Smalltalk prototype. Encouraging results are obtained.