This dissertation describes research into image processing techniques that enhance military operational and support activities. The research extends existing work on image registration by introducing a novel method that exploits local correlations to improve the performance of projection-based image registration algorithms. The algorithm is shown to operate in low signal-to-noise ratio (SNR) conditions and to significantly improve registration performance by as much as a factor of 5.5 in mean-squared error over existing projection-based registration algorithms at a minimal computational cost.

The dissertation also extends the bounds on image registration performance for both projection-based and full-frame image registration algorithms and extends the Barankin Bound from the one-dimensional case to the problem of two-dimensional image registration. The Cramer-Rao and Barankin bounds are calculated for registration performed using 2-D registration algorithms and compared to bounds on registration estimates calculated using computationally efficient projection-based registration algorithms. It is demonstrated that in some instances, the Cramer-Rao lower bound is an overly-optimistic predictor of image registration performance and that under some conditions the Barankin bound is a better predictor of shift estimator performance. These conditions include low-SNR imaging and imaging under defocus error, two conditions which are frequently encountered in military imaging systems that employ passive infrared, light radar (LIDAR), and synthetic aperture radar (SAR).

The research looks at the related problem of single-frame image denoising using block-based methods. The research introduces three new algorithms for single-frame image denoising that operate by identifying regions of interest within a noise-corrupted image and then generating noise free estimates of the regions as averages of similar regions in the image. The algorithms are shown to outperform Wiener and median filtering over a wide range of noise conditions but are most effective in images with very low signal-to-noise ratios.