Subpixel registration and super-resolution techniques for digital image sequence restoration

Many imaging sensor systems obtain a sequence of slightly misregistered images of a static (or slowly changing) scene due to relative motion between the scene and the camera. Such an image sequence can be used to synthesize an image of higher resolution and less degradation than provided by any of the original image frames. One method of utilizing this image sequence is to decompose the problem into two parts: first register the sequence of images to subpixel accuracy, then obtain the restored image to estimate the true scene intensities from the higher resolution image obtained by merging the frames together. This research focuses on the development and experimental testing of algorithms to solve these two problems.; Two images are used. One is the classical stochastic model based on Gaussian random fields. The other one is a modification of a deterministic model which uses two dimensional B-splines.; These images are used to derive two different approaches to the subpixel image registration problem. In both cases a functional of the two images is computed and the estimate of the displacement between the two image frames is found as the location of the maximum of the functional. The functionals both contain a filtered cross correlation term in which the filter is lowpass in nature and depends on the value of the estimated displacement. Experimental results are reported for performance on real image data. Both algorithms provide true subpixel registration performance.; A new approach is taken to the formulation of the constrained restoration, or super-resolution, problem. It is shown that many previous formulations of the constrained restoration problem can be put into the context of state estimation theory, and new results are established which guarantee that the scene may be estimated from the sequence of images under more general conditions than the previous theory on super-resolution would allow. A new super-resolution image restoration algorithm is developed based on a time sequential Kalman estimator. A suboptimal but computationally simpler version of the algorithm is proposed for processing two dimensional images. The algorithms are applied to both synthetic and real image data to demonstrate their validity.