This paper presents a technique that performs multi-frame super-resolution of differently exposed images. The method first employs a coarse-to-fine image registration method to align image in both spatial and range domain. Then an image fusion method based on the maximum a posterior (MAP) is used to reconstruct a high-resolution image. The MAP cost function includes a data fidelity term and a regularized term. The data fidelity term is in the L2 norm, and the regularized term employs Huber-Markov prior which can reduce the noise and artifacts while reserving image edges. In order to reduce the influence of registration errors, the high-resolution image estimate and registration parameters are refined alternatively by minimizing the cost function. Experiments with synthetic and real images show that the photometric registration reduce the grid-like artifacts in the reconstructed high-resolution image, and the proposed multi-frame super resolution method has a better performance than the interpolation-based method with lower RMSE and less artifacts.
This paper describes an approach to reconstructing wavefronts on finer grid using the frozen flow hypothesis (FFH), which exploits spatial and temporal correlations between consecutive wavefront sensor (WFS) frames. Under the assumption of FFH, slope data from WFS can be connected to a finer, composite slope grid using translation and down sampling, and elements in transformation matrices are determined by wind information. Frames of slopes are then combined and slopes on finer grid are reconstructed by solving a sparse, large-scale, ill-posed least squares problem. By using reconstructed finer slope data and adopting Fried geometry of WFS, high-resolution wavefronts are then reconstructed. The results show that this method is robust even with detector noise and wind information inaccuracy, and under bad seeing conditions, high-frequency information in wavefronts can be recovered more accurately compared with when correlations in WFS frames are ignored.