EM and MAP based super-resolution and parameter estimation

In this thesis digital image Super-Resolution techniques using the Expectation Maximization (EM) and Maximum a posteriori (MAP) techniques are discussed. A Super-resolution algorithm using the EM algorithm in the frequency domain taking advantage of the diagonalization property of the DFT is discussed. Super-resolution using a combination of EM and MAP algorithms is also presented.; We have considered the problem of obtaining a high resolution (HR) image from multiple low resolution (LR) frames. For this approach, the low resolution frames are required to have sub-pixel shifts among them, thus making each low resolution frame provide unique information about the scene which can be used to construct a high resolution frame. A method for interlacing the multiple LR frames to form a high resolution grid is discussed. This approach reduces the complexity of the super-resolution algorithm and provides for efficient implementation in the frequency domain.; Point spread function (PSF) estimation, which is critical to the quality of the final image is also discussed. The practical scenario in which the PSF is partially known is considered. The image observation model accounts for blurring due to PSF and subsampling. In addition, the LR frames are assumed to be corrupted by additive white Gaussian noise (AWGN) whose variance is estimated by the algorithm.; The super-resolution algorithm is applied to synthetic data as well as real data. Given a set of LR frames the algorithm forms a HR grid and performs image restoration on the HR grid to yield an image with higher pixel density. Experimental results are presented and conclusions are drawn.