A Multilevel Approach to Sequential Detection of Pictorial Features

This paper considers the problem of detecting the local similarity between templates in a given class and a given image using a hierarchically ordered sequential decision rule. When the given set consists of a large number of templates and the number of locations in the image matching any of the templates is small, it is wasteful to examine each of the templates at every location in the image for a match. Instead, it is proposed that the set of templates be partitioned and a ``representative template´´ be defined for each of the partitions. Several levels of partitioning are defined. Elimination of mismatching locations and termination of computation can take place at each, level of detection. Each level of testing is over a more restrictive subset of the template class than the previous level. The paper presents a general formulation of this approach and gives criteria for selecting representative templates, the ordering of components of a template vector for error evaluation, and the threshold sequences to be used in deciding about a ``match.´´ Suboptimal solutions are given satisfying these criteria. Illustrative examples are provided showing recognition of linear features in test patterns and photographs obtained by aerial and spaceborne sensors.