Image super-resolution: Iterative multiframe algorithms and training of a nonlinear vector quantizer

Images acquired by ground-based telescopes are severely degraded by atmospheric turbulence effects. New algorithms are presented for restoration with super-resolution of satellite object images from sequences of turbulence-degraded observations. Super-resolution refers to recovery of Fourier spectral components outside the optical system passband. Modern wave front sensor (WFS) can measure the optical distortions caused by the atmosphere. Such measurement can be used for (1) control of an adaptive optics (AO) system; (2) for post-processing of the uncompensated image; and (3) for a hybrid approach involving partially compensated images. This study focuses on the second of these approaches. Quantitative simulation of imaging through turbulence and WFS are used to demonstrate the performance of new super-resolving multiframe algorithms based on Bayes maximum a posteriori (MAP) criterion. The original and object images are assumed to have Poisson statistics. The resulting Poisson MAP algorithms extend the single frame version to the multiframe case. Super-resolution is demonstrated for realistic conditions.; In the blind deconvolution problem, both the original image and the degradations must be derived simultaneously from the recorded images without the aid of WFS. We investigate this problem and propose a new multiframe algorithm based on Bayes maximum likelihood. Strict constraints such as positivity and finite bandwidth are incorporated using nonlinear reparameterizations. Nonlinear conjugate gradient techniques are employed along with implementation on the massively parallel IBM SP2, in order to meet the computational demands of these algorithms. Super-resolution is demonstrated for realistic circumstances.; On a related subject, nonlinear interpolative vector quantization (NLIVQ) is presented as a tool for the novel application of vector quantization (VQ) to super-resolution of diffraction-limited images. The algorithm is trained on a large set of image pairs, consisting of an original and its diffraction-limited counterpart, and exploits the statistical dependence between blocks of pixels in the two images. The discrete cosine transform (DCT) is used to manage the codebook complexity and simplify training. Simulation results are presented which demonstrate improvements in the visual quality and peak signal-to-noise ratio. A study of restored image spectra reveals modest super-resolution. The prospects for this technique are promising.