In optical imaging, image quality is not only determined by the system itself but also by the media in which light traverse. Differences in the refraction index of the media encountered by a light wavefront produces phase aberrations which distort the image received from the original object. Currently, there are two main approaches for solving this problem: adaptive optics, which rely on deformable mirrors and wavefront sensors for correcting the phase aberration before it reaches the imaging sensor; and post-processing techniques, which try to estimate the object after receiving distorted images.

MFBD methods are a family of algorithms that are capable of reconstructing the object by fusing the information carried by a set of differently aberrated images. These techniques are widely used in current optical systems, allowing a notable increase in image quality in most situations. However, there are cases in which they are not applicable; for example, looking at a dynamic object (e.g., a bird flying) or looking through static aberrations (e.g., in microscopy applications); but certain modifications in the optical system can be used for solving this problem.

Actual optical devices have only one aperture, thus creating one full-size image on the imaging sensor but, by segmenting the pupil, several images can be retrieved at the same time with different aberrations (i.e., light follows a distinct path for each aperture). However, using a multi-aperture system implies that there is less imaging sensor area available for each aperture, thus obtaining images with less resolution. Nonetheless, MFBD algorithms can usually be extended in order to support SR, a technique that allows the increase of the object resolution by retrieving extra information from the displacement between images.

This thesis is focused on the development of a functional prototype of a multi-aperture optical system that can do real-time object reconstruction. As a MFBD technique is needed, the novel TIP algorithm (developed at Delft Center for Systems and Control) is selected. In order to achieve a fast and reliable reconstruction, the algorithm is: modified for increasing its robustness against noise, expanded in order to support SR and implemented efficiently in both CPU and GPU. Finally, the system is tested in a real environment, showing promising results.