Storage-efficient quasi-Newton algorithms for image super-resolution

Multiframe image super-resolution algorithms can be used to obtain a higher-resolution higher-quality image from a set of low-resolution, blurred, and noisy images. Very often, these algorithms rely on an optimization-based inversion of the image acquisition model. Recently, two algorithms for grayscale and hybrid demosaicing and color super-resolution have been proposed by Farsiu et al. These algorithms are of practical interest because they are fast and also they can overcome mismatches in the assumed acquisition model. However, they rely on the use of steepest-descent minimization which is inefficient in highly nonlinear and ill-conditioned problems like super-resolution. In this paper, we use two storage-efficient quasi-Newton algorithms, the memoryless and the limited-memory BFGS algorithms, to improve the performance of the super-resolution approaches proposed by Farsiu et al.