Shape Recognition by Computer in Simulated Aerial Images

摘要

Visual recognition of objects by a machine involves classifying an input usingknowledge about the kinds of objects expected in the domain. Model-based systems maintain a knowledge of objects in the domain in the form of a representation which can be compared to the unknown input. Since a given object type may appear in a variety of forms and under a variety of viewing conditions some efficient yet flexible means of guiding the recognition process to consider and then verify the object identity is necessary. The utility of low-resolution shape information to constrain object recognition was investigated in the context of a system which is predicated upon a component description of objects. A computationally intensive prepass using a syntax for combining components yields a universe of constructions which are coded into a construction relation feature (CRF) map. Each construction is coded into the N-dimensional map according to a shape parameterization of its low-resolution image (each dimension codes a shape feature). From a subset of these constructions the object models are specified in terms of their component structure. The CRF map thus links the low resolution shapes of instances of an object to its object model. To recognize an unknown object the input is first converted to low resolution. Then, shape parameters are taken (for example, in terms of its relative elongation and compactness).