Wavelet-Based Bayesian Methods for Image Analysis and Automatic Target Recognition

摘要

This work investigates the use or Bayesian multiscale techniques for image analysis and automatic target recognition. We have developed two new techniques. First, we have develop a wavelet-based approach to image restoration and deconvolution problems using Bayesian image models and an alternating- maximation method. Second, we have developed a wavelet-based framework for target modeling and recognition that we call TEMPLAR (TEMPlate Learning from Atomic Representations) . TEMPLAR is can he used to automatically extract low- dimensional wavelet representations (or templates) or target objects from observation data, providing robust and computationally efficient target classifiers. On a more theoretical level, we have developed a framework for multiresolution analysis or likelihood functions, which extends wavelet-like analysis to a wide class or non-Gaussian processes. In another line of investigation, we are exploring a new imaging application known as network tomography. The goal of this work is to characterize the internal performance of communication networks based only on external measurements at the edge (sources and receivers) of the network. In the coming year, we plan to focus on four key research areas. First, we will develop theoretical hounds on the performance of multiscale/wavelet estimators in non-Gaussian environments including Poisson imaging applications. Second, we will study the use of complex wavelets in image restoration and target recognition problems. Third, we will develop automatic methods for segmenting imagery (SAR, FLIR, LADAR) based on complexity- regularization methods. Fourth, we will continue to develop a unified framework for communication network tomography and investigate new tools for network performance visualization.