Recognition performance from synthetic aperture radar imagery subject to system resource constraints

The problem of automatic target recognition (ATR) can stated be as the problem of inferring, from the output of one or more sensors directed at a scene, the classes to which objects in the scene belong and the properties of those objects such as sub-class, pose, and states of articulation. We consider the specific problem of ATR based upon synthetic aperture radar (SAR) imagery, though the principles employed are applicable in the wider context of object recognition. Approaches to automated recognition are developed in the context of a communication-based model. The recognition system is viewed as a recipient of information from two sources: a scene containing the object to be recognized and a database characterizing the objects to be recognized. The overall accuracy of the system is dependent upon the properties of the scene and sensor, the accuracy of the imaging model on which the system is based, and the accuracy of approximations made for the purpose of system implementation. These last two items have a direct impact on the computational resource requirements of a recognition system. The accuracy of a system is thus directly related to the available resources, such as the number of processor cycles used, mass storage requirements, network bandwidth utilization, elapsed time, etc. This relationship can be characterized by an accuracy-consumption curve which is useful for comparing alternate approaches to recognition and for exploring the space of system design possibilities.; A statistical hypothesis testing approach is followed and several variants of four probabilistic models for SAR imagery are discussed. A methodology for assessing the validity of model assumptions is developed which accommodates large numbers of small samples with unknown distribution parameters. This methodology is applied to assess the SAR models using sample SAR data. Based on the assumed models, algorithms for estimating model parameters from training data and for inferring the class and pose of objects in SAR imagery are presented. Analytical expressions for the probability of error in a binary hypothesis testing problem are derived. Several methods are considered for declining to classify objects which are not represented among a database of known objects. These methods of so-called confuser rejection are based on estimated measures of relative information, tests of significance, and Bayes minimum risk, respectively.; The ATR algorithms are applied to actual SAR data under a wide range of approximations governing the infinite variety of target pose. The degree of approximation determines both the recognition accuracy and the system cost in terms of model storage, communication, and processing. Each approach is characterized in terms of the lowest achievable error rate as a function of system resource consumption over the set of approximations considered. Object models are considered which are successively-refinable in pose, that is can be incrementally refined from coarse representations. Such models allow likelihood functions to be embedded into a tree structure where nodes in the resolution tree represent successively smaller pose ambiguity. Maximization of likelihood functions becomes a search of the resolution tree, and the tree structure can be exploited to locate good possibilities quickly. This approach leads directly to successively-refinable decisions in which an initial classification is gradually refined as more branches of the resolution tree are explored. Successively-refinable decisions allow the accuracy-consumption properties of an ATR system to be adjusted dynamically, terminating the search when a pre-specified quantity of resources has been expended.