The objective of this project is to develop algorithms that will enable intelligent imaging add vision systems to operate robustly and rapidly with very noisy sensor data. Current methods perform poorly in this regime due to a lack of well-defined optimality properties, `brittleness' of the underlying discrete representations, and a deluge of processing inconsistencies and noise artifacts. The improved performance is achieved by a cooperative solution of low-level signal processing and high- level object recognition tasks subject to numeric and symbolic constraints, instead of the classical sequential processing method. The overall approach is Bayesian, and is formulated as the minimization of an energy surface. All the image-processing operations result automatically as by products of a gradient- based optimization process. This includes the synergistic interaction of low and high-level variables, satisfaction of pattern constraints, and feedback of high-level modelling information. From an architectural standpoint, the proposed method offers a valuable uniform strategy for parallel computation, which is duly exploited. This work is expected to fundamentally impact applications that are required to operate with noisy sensor data in a limited time frame. Encouraging preliminary results have been computed on a DAP massively- parallel processor.
