A scene similarity metric for matching configurations of image objects

This paper proposes a novel solution to querying image databases by matching raster features to imagery completely in the raster/spatial domain using the shape of single features together with their spatial relations as the matching primitives. The core feature-based matching module combines a revised least-squares matching algorithm, to accomplish the matching process on binary images, with a unique implementation of a feature library that organizes and links query objects with their images in the database, thus enabling fast and efficient real-time retrieval of relevant imagery. The purpose is to extend this previous work of ours, which matches on the shape of single features only, to include querying configurations of image objects. As such, both spatial reasoning, e.g. the topological (disjoint, touching, overlapping, etc.), directional (north, south, east, etc.), and metric (distance) relationships between individual features, together with feature shape matching are requirements for proper image retrieval. More specifically, we will define a methodology to evaluate similarity between a user-defined query scene and the retrieved images. To do this, it is required to define a function Smet that assesses the similarity metric between a query configuration Q and an image I in the database. Such a function will combine different similarity functions for individual object shapes, and the spatial relations between them. The result of this investigation will be a proposal for solving the problem of scene similarity querying in image databases that will provide the basis for future work on implementation and act as the framework for further extensions into temporal reasoning.