Heuristics for Image Retrieval Using Spatial Configurations

This paper describes a strategy for content-based image retrieval via spatial configuration of objects within an image. The proposed strategy locates “pattern” images within “text” images, where the objects in each image are specified only in terms of their pairwise spatial relationships. Thus both pattern and text are represented by square matrices whose elements are drawn from an “alphabet” of spatial relationships. A match is found when the pattern image is identified as a “substructure” of a given text image. Experimental results relating to retrieval of images with low image content are shown. The approach can be used in situations where methods involving features such as colours and texture are inappropriate. We believe that such algorithms would find application in topological/GIS applications where the aerial/satellite images are very similar in terms of low-level image content.