Introduction

One principal goal is to pursue an integrated set of research problems associated with automatic segmentation, detection and recognition of object/target (AOD/R), spanning the complete processing chain from sensor signal processing to image analysis to object recognition, thereby allowing us to understand how these individual processing stages interact and to define new approaches. By considering the entire processing chain, we establish a critical audit trail so that performance characteristics of one stage of the processing can be used as models for the information provided as input to the next stage in the chain. In this way fundamental limits on the quality of information-e.g., on the extraction of particular features-can be propagated forward, ensuring that the design of subsequent stages are based on realistic expectations about the information that is available. Conversely, performance requirements at each processing stage can also be reflected back to the quality of the input to that stage and hence to requirements for the preceding stage. Our ultimate goal is to provide a rational framework for determining absolute performance limits, for designing systems that achieve them, and for suggesting how performance might be improved. For example, if object recognition analysis identifies salient features that would considerably enhance performance, an integrated framework could then guide the quantification of sensing enhancements required to achieve the desired feature extraction. In addition, by considering the full AOD/R chain, we can identify suboptimalities in this decomposition that would be overcome if a different cut through the processing structure were taken-e.g., the integration of recognition functions, such as pose estimation, into the preceding stage of synthetic aperture radar (SAR) image formation. To realize these cross-disciplinary objectives we are taking an approach that has two important components: the use of statistical models and the development of multiresolution methods. Statistical models provide an extremely flexible framework in which to quantify information and uncertainty and, in particular, to realize the bi-directional audit trail mentioned previously. The use of multiresolution methods, ranging from wavelet transforms to statistically optimal multiresolution algorithms, offers not only the promise of computationally superior algorithms but also an explicit and rational basis for addressing resolution/accuracy trade-offs. In addition, using multiresolution representations for the data, the scene, and the objects leads directly to methods for multiresolution image formation, for the extraction of multiresolution geometric or fractal features, and for multiresolution object recognition. We are currently at the early stages of building both this multiresolution statistical framework and the interdisciplinary links that are crucial for integrated AOD/R. The modeling and analysis of particular types of data-primarily multimodal medical data sets (MRI, PET, CT) and and visual information processsing --provide the vehicle for our current efforts. (Understanding vision through Psychophysics, Electrophysiology, Neuropsychology, functional MRI, and Computational Modeling.)