A unified approach to the superresolution of optical and synthetic aperture radar images

This work addresses the issue of superresolving optical and synthetic aperture radar (SAR) images. Superresolution is the process of obtaining an image at a resolution higher than that afforded by the sensor used in the imaging. The concept of signal resolution is shown to be intimately related to the notion of perfect reconstruction. As such, the major issue addressed in this work is in establishing an appropriate set of bases for the reconstruction of images in these domains in terms of a finite set of collected samples.; The existing theoretical foundations for perfect reconstruction in the aforementioned domains is expressed in terms of linear projections and infinite extent data. However, practical restrictions leading to finite collected data limits the resolution afforded by the theoretically established bases, i.e. the sinc bases in the spatial domain and the Fourier bases in the frequency domain. Superresolution deals with this issue by incorporating a priori information into the process of establishing an appropriate set of projections for reconstruction.; In the optical domain, a priori information is extracted in an unsupervised manner from low and high resolution versions of digital images of the same scene. This information is provided by the local neighborhood about the point to reconstruct. This is in agreement with the local weighting of the theoretically established sinc bases for perfect reconstruction. In the SAR domain, a priori assumed models are inherent to the reconstruction. The information is provided globally in terms of a best match to these models. This is also in agreement with the global weighting of the theoretically established Fourier bases for perfect frequency reconstruction.; The superresolution of images is accomplished by means of a modular architecture using a two-step approach. First, the collected samples are compared with a priori determined clusters for each image point to reconstruct. Second, each cluster is linked to its optimal reconstructor which was determined from the collected set of samples. The results obtained herein are compared against several accepted sets of bases. Optical images of real scenes as well as textures are tested. SAR data from real metallic polyhedral objects as well as U.S. military tanks are imaged and tested for automatic target detection performance. It is shown that by appropriate incorporation of a priori information into the superresolution procedure, better reconstructed images result.