| Image superresolution refers to image processing algorithms which produce high quality, high-resolution images from a set of low quality, low-resolution images. In many visual applications, both civilian and military, the imaging sensors have poor resolution outputs. When resolution can not be improved by replacing sensors, either because of cost or hardware physical limits, we can resort to superresolution algorithms. Even when superior equipment is available, superresolution algorithms are an inexpensive alternative.; Superresolution is a computationally intensive process. Some video applications may require superresolution to be done on-the-fly; data must be processed as they are received. To that end, it must take advantage of inherent regularity and structure of the problem. Superresolution algorithms must be robust with respect to various sources of image degradations. These include unknown sensor noise, unknown or varying camera characteristics from frame to frame, unknown modelling error, etc. Furthermore, the algorithms must be driven only by the sensor data. No detailed information about noise or camera characteristics is given.; The goal of this thesis is a complete superresolution algorithm which can be applied in realistic applications. The thesis examines superresolution under two different frameworks: an iterative approach and an interpolation-restoration approach. Fast and robust techniques are presented for various components of superresolution. We develop a projection-based framework for frame-to-frame motion estimation and image registration aspects of superresolution. For the iterative approach, the high resolution image estimate is the solution to a regularized least squares system. We propose new preconditioners to accelerate convergence for the conjugate gradient method applied to the regularized least squares problem. We also consider the issue of regularizing discrete ill-posed underdetermined problems and derive new formulas for two regularization parameter estimation techniques. Very often, the blurring operators are unknown or reliable estimates are unavailable. We develop new techniques to identify unknown blur from multiple low resolution frames. We also present a novel wavelet interpolation-restoration approach to superresolution. Numerical experimental results demonstrate the effectiveness of the proposed methods. |
